https://doi.or g/10.31449/inf.v49i7.6615 Informatica 49 (2025) 33–48 33 Deep Learning-Driven Edge-Enabled Serverless A r chitectur es for Animal Emotion Detection Shajulin Benedict 1, 2 , Rubiya Subair 1 1 Department of Computer Science and Engineering, Indian Institute of Information T echnology Kottayam, Kerala, India 2 T echnical University Munich, Garching, Germany E-mail: shajulin@iiitkottayam.ac.in, shajulinbenedict@mytum.de, rubiyasubair .23phd21001@iiitkottayam.ac.in Keywords: Animal emotion, deep learning, edge intelligence, serverless Received: July 9, 2024 Animal emotion detection, including elephant emotions, is highly possible, but what the traditional emo- tion detection appr oaches highlight is their blatant ignorance of adopting edge-enabled intelligence and serverless-based solutions, both of which ar e affor dable. T r eating the emotions of animals incr eases their pr oductivity , especially among trained elephants when subjected to carrying logs or undertaking gar gan- tuan tasks. However , existing infrastructur es ar e inefficient in handling long-running animal emotion detection-r elated tasks. This article pr oposes a deep learning-driven edge-enabled serverless ar chitec- tur e after evaluating several existing animal emotion detection techniques. Additionally , we perform an exploratory study on the cost impact of incorporating serverless-enabled appr oaches to animal emotion detection ar chitectur es. W e observed that the pr oposed edge-enabled serverless ar chitectur es saved over 13,000 dollars annually compar ed to traditional animal emotion detection appr oaches. In addition, the ar - ticle pr ovided a few r esear ch dir ections to develop novel edge-enabled serverless ar chitectur es that boost socio-economic situations while avoiding human-animal conflicts. Povzetek: Pr edlagan je sistem za zaznavanje čustev živali na osnovi globokega učenja. Model povečuje učinkovitost analize, zmanjšuje str oške in izboljšuje interakcije človek-žival. 1 Intr oduction Many animals, the vital sources of enabling a sustainable natural environment, have been integral companions of hu- mans for years to share emotions. Emotions are, in general, considered distinct physiological responses that trigger ac- tivation signals from various sources, such as the face, body movements, and eye/ear/tail directions. The emotional traits of animals can bring forth immense psychological relations, leading to research directions that relate to natural survival ethos. For instance, the postures of animals, while proactively reacting to natural disasters, prevent major losses in lives, especially in rural areas. Re- searchers have observed an anomaly in the behavior of toads and birds before earthquakes. Also, monitoring the emotions of farm animals such as pigs, cows, and sheep, has increased productivity in yielding milk or food prod- ucts. These prior experiences have led researchers to cor - relate the emotions of animals with productivity measures. In fact, emotions in animals have driven researchers to pursue research in various dimensions. For instance, i) al- most many veterinary researchers have reckoned in sum- ming up the animal health behaviors to study human health issues – i.e., the treatment of rare psychological human dis- eases has been analogously studied using laboratory ani- mals [ 48 ]; ii) researchers have focused on monitoring the emotions of trained wild animals, such as elephants, that are involved in carrying logs or erecting poles in remote loca- tions or forests to increase the much-awaited human-animal bonds [ 53 ]. Although many animals exhibit emotion, there are sev- eral challenges in identifying and classifying them appro- priately . The most notable challenges are listed as follows: 1. V arying Degree of Emotions – Animals have dif fering degrees of emotions. They reveal sadness and happi- ness in varying degrees when compared to most hu- man beings around the globe [ 5 ]. 2. Dif fering Emotions among Animals – The empathy re- vealed in rats is dif ferent from the empathy exhibited by elephants. Researchers have recorded moments that portray the dif ferences among animals. For in- stance, elephants mourn in the worst situations, such as the deaths of mahouts. On the contrary , dogs com- fort their masters while strange incidents happen at home. 3. V arying Emotional Features – The features associated with emotions are unique to animals. Hence, develop- ing one robust AI-assisted solution or framework that suits detecting the emotions of all animals is a chal- lenge. 34 Informatica 49 (2025) 33–48 S. Benedict et al. 4. Algorithm Designs – T ens of thousands of learning al- gorithms exist on the market to detect emotions in hu- man faces or mobile-assisted applications [ 35 ]. How- ever , for developers, there are not many learning al- gorithms that could accurately capture the emotions of animals with limited computational or communication requirements. There have been ef forts in the past to study the emotions of animals [ 57 ], [ 68 ] – for instance, researchers have de- veloped sensor -enabled systems and AI-assisted solutions to detect the emotions of animals. However , it is crucial to understand the pros and cons of these solutions so that new approaches can be designed or developed. This article examines the existing emotion intelligence methods used by researchers to detect animal emtions. The article highlights the importance of incorporating edge- enabled and serverless-oriented solutions when designing animal emotion detection frameworks or architectures. The article delves into the importance of integrating serverless approaches into architectures by examining the costs as- sociated with using computational resources and memory components for animal emotion detection. Additionally , it throws light on future novel research directions and ap- proaches to animal emotion detection frameworks. The ar - ticle contributes to classifying animal emotion detection- related works and delivers a taxonomy of animal emotion- related research works. The rest of this article is described as follows: Sec- tion 2 provides a taxonomy of emotion detection ap- proaches in animals; Section 3 expresses the necessity of edge nodes and edge-enabled serverless-based emotion de- tection frameworks with suitable cost-based exploratory il- lustrations; Section 4 examines a few possible research di- rections that a few computational researchers could under - take shortly; and Section 5 provides a few conclusions of the article. 2 Animal emotion detection – A taxonomy T raditionally , animal emotions were manually detected. For instance, mahouts had to keenly observe the elephants’ movements, facial expressions, and sound characteristics before performing actions. In recent years, the manual ap- proach has been replaced with IoT -assisted solutions or AI- based intelligent solutions [ 75 ]. This section explains two broad classifications of ani- mal emotion detection mechanisms – Invasive and Non- invasive methods. Before delving into these methods, we have emphasized the dif ference between animal and human emotions. Additionally , we have listed animals that are of- ten utilized in the state-of-the-art literature to detect emo- tions in various contexts. 2.1 Human emotions vs. animal emotions Often, humans deliver emotions using multiple modes of linguistic communication, such as texts or vocals, apart from facial/body expressions or movements. The emo- tional states of humans have been widely discussed in the past by several researchers. Notably , [ 64 ] have developed emotion-oriented facial datasets for the Indian community . Similarly , facial emotion datasets have been developed in the past to be applied in several applications such as driver assistance, fraud detection, culprit detection, and so forth. In the past, the majority of researchers classified human emotions into six categories. According to [ 22 ], human emotions are classified as: i) happiness, ii) sadness, iii) fear , iv) disgust, v) anger , and vi) surprise. However , recently , a few authors [ 18 ] have observed a few more unique emo- tional states in humans. Accordingly , human emotions are represented in 27 dif ferent states, such as admiration, ado- ration, aesthetic appreciation, amusement, anger , anxiety , awe, awkwardness, boredom, calmness, confusion, crav- ing, disgust, empathic pain, entrancement, excitement, fear , horror , interest, joy , nostalgia, relief, romance, sadness, sat- isfaction, sexual desire, and surprise [ 18 ]. Pythagoras and Charles Darwin mentioned that the emo- tions of animals are equivalent to human emotions [ 5 ]. But many other researchers have recently expressed that the emotions of humans are quite deeply intertwinkled in most cases when compared to animals – i.e., the emotions of hu- mans have a mixed expression in some situations that the animals could hardly express. Hence, excerpts conclude that animal emotions dif fer from human emotions. In [ 55 ], the authors have classified the core emotions of animals into seven categories: seeking, fear , rage, lust, care, panic, and play . These emotions dif fer depending on the type of animal. For instance, authors of [ 51 ] have ex- plored the emotions of elephants in detail; the authors of [ 42 ] have translated human emotion traits to study the emo- tions of animals, with a specific focus on handling concep- tual emotions. In general, elephants reveal emotions in several situa- tions. The most commonly observed incidents in which ele- phants showcase emotions are listed below: a) Joy during birth – i.e., elephants trumpet and run around each other to express joy at birth; b) Grief during death – i.e., ele- phants mourn if death occurs among loved ones. During grief situations, elephants reiterate the incidents due to their long memory power . Elephants exhibit certain characteris- tics that are exceptional to many other animals; c) Anger Emotion – elephants act angrily in several situations, es- pecially when their habitats are occupied by humans. Simi- larly , they express angry emotions when mahouts ur ge them to do lar ger tasks without suf ficient food or enough rest. T ypically , the spread ears, V -ear , and distracted working style of elephants indicate a sign of threatening people or environments [ 37 ]; and, d) Empathy Emotion – elephants also exhibit empathy on three major occasions: i) consoling people or other elephants during any death occurrences, ii) Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 35 Figure 1: Emotion detection of animals – A taxonomy defending or protecting young elephants from other preda- tors, and iii) stopping fights, especially when their loved ones are forced to fight each other during unavoidable cir - cumstances. 2.2 Animals involved Animals involved in observing emotions dif fer based on the location where they are monitored. Accordingly , they are classified as follows: 1. Farm animals – Farm animals, such as dairy cows, pigs, hens, horses, and so forth, are animals utilized for agricultural purposes such as yielding milk, lay- ing eggs, and providing nutrients. Understanding the emotions of these farm animals and providing better shelter and food can increase the economic capabilities of farmers. It can indirectly increase the financial posi- tion of the livestock sector of a nation. In the past, the emotions of dairy cows have been observed in pastures [ 4 , 51 ]; pigs have been monitored in farms [ 13 , 51 ]; and, the emotions of sheep have been assessed in farms [ 45 ]. 2. Pet animals – Pet animals are often utilized by hu- mans to provide a better companion in their lives. The utilization of pet animals such as cats and dogs has increased in multitudes in recent years owing to the loneliness of human inhabitants that they experienced during the COVID-19 era. It is manifested that the productivity of working professionals has increased in multitudes due to the association of pet animals. 3. Laboratory Animals – Laboratory animals are in- volved in studying the relevant features of humans or their health conditions. The most commonly utilized animals for laboratory-based study purposes are rats and mammals. Learning the emotions of these animals before dissecting them is considered a crucial task for several veterinarians or animal researchers [ 1 ]. In some situations, animals are experimented on in test rooms rather than in laboratories. This is a specific location where emotions are experimented with us- ing sophisticated measurement devices or sensor units. However , the test rooms are not utilized for students or for grading them. For instance, a few pigs have been monitored by researchers in a test room consisting of four positioned cameras associated with computer vi- sion image analyzing systems to study their behavior [ 68 ]; also, the emotions of dogs have been evaluated in a room by monitoring their facial expressions and providing rewards to them [ 9 ]. 4. W ild Animals – W ild animals such as elephants [ 65 ] [ 37 ], lions, tigers, or so forth that live in forests, such as sparsely dense forests or heavily dense forests, might expose varying degrees of emotion. For in- stance, elephants’ vocalization with varying degrees of emotions has been recorded in forests to capture the wild behavior of elephants [ 70 ], [ 76 ], and [ 16 ]. The purpose of characterizing and learning such emotions is to assess the patterns of invading wild elephants or to save tribal communities from possible elephant at- tacks. In [ 58 ], authors have studied the characteristics of 36 Informatica 49 (2025) 33–48 S. Benedict et al. sheep when they wandered in a wild forest. This ex- ercise has been carried out to observe the diversity of data sources – wild vs. domestic impacts. 5. T rained W ild Animals – A few wild animals are trained to undertake gar gantuan tasks such as carrying logs or showcasing performances in the circus arena. W ild animals performing in circus arenas, such as ele- phants, tigers, and lions, express emotions considering the trainers’ reactions or audiences’ appreciation lev- els. 2.3 Emotion-r elated featur es The emotions of animals are exhibited in animals through specific features. The core features that are instrumental in identifying the emotions of animals are listed as follows: 1. Pig Emotion Identification Featur es – The emotions of pigs are evaluated based on the angle of their ears, snout ratios, and eye movements. If pigs exhibit ag- gression, ears move forward with a minimal snout ra- tio; similarly , their eyes are narrowed during retreats [ 13 ]. 2. Cow Emotion Identification Featur es – The emotions of cows are identified based on the movements of their eyes and ears. If they are excited, the ears are wide open and the eye positions are broadened. 3. Elephant Emotion Identification Featur es – Elephants are considered the lar gest land mammal on the planet, having the lar gest brain that weighs up to 5.5 kg. It has a huge physical structure of 3 to 4 meters tall and weighs from 2 to 7 tons. The trunks have over 1 mil- lion muscles to undertake several gar gantuan tasks, such as carrying logs or removing trees. The trunk portion of the elephant is utilized for several activi- ties, such as defending, breathing, feeding, gathering, smelling, drinking, lifting, and sensing. The brain of elephants is so peculiar due to their enduring long-term memories, which enables them to remember mahouts or emotional incidents even after long years. There are two types of elephants: a) Asian Elephants : Asian elephants are most commonly found in India, Sri Lanka, and Sumatra. These elephants have twin- domed heads compared to African elephants. They weigh heavier than African elephants in most cases. b) African Elephants : African elephants are of two types – bush elephants and forest elephants. The bush elephants are found in several parts of Africa. They are gigantic in nature when compared to the other elephants – i.e., they weigh over 10 tons. On the contrary , the forest elephants are smaller and have rounded ears. The poaching and defending character - istics dif fer among these two species of African ele- phants. 4. Dog Emotion Identification Featur es – Several re- search studies have been carried out to study the emo- tions of dogs in the past. The crucial features of dogs that impact their emotions are blinking eyes, flattening ears, moving lips, licking noses, and so forth [ 17 ]. 5. Horse Emotion Identification Featur es – Learning the emotions of horses is based on the evaluation of their eyes, ears, nose, and neck positions. Depending on dif ferent emotional traits, horses position their body parts, exhibiting unique emotional features. For in- stance, i) open eyes, stif fly forward ears, open nos- trils, parallel necks, and higher head positions are the indications when horses are alarmed [ 17 , 63 ]; ii) open eyes, stif fly backward ears, and slightly closed nos- trils, are the indications when they are annoyed; iii) open eyes, pointed forward ears (mostly relaxed), and an open mouth with the relaxed neck are specific fea- tures when they are curious about pursuing any tasks; iv) shut eyes, pointed sidely ears, relaxed mouth, and parallel neck positions are signs that indicate a relaxed horse; v) rotated ears, dilated nostrils, and raised chin are the indications for revealing pains by horses [ 59 ]. 2.4 Emotion detection methods There are two broad categories of emotion detections in an- imals – i) Invasive and ii) Non-Invasive. In the invasive method, sensors or electronic-based emotion detection de- vices are either permanently or temporarily inserted into the bodies of animals. These electronic devices can constantly upload measurable properties such as heart rate, tempera- ture, glucocorticoid levels, or physiological changes to the connected cloud or fog-based compute services using wire- less communications, including 4G/5G mobile networks for evaluating the emotions. Additionally , animals are uti- lized to detect natural disasters using invasive techniques [ 32 ]. Although this approach has been widely applied in hu- man emotion detection methods, it is not well discussed or practiced in animal research, particularly when considering small-sized animals such as cats, rats, and so forth. How- ever , elephants and similar kinds of endangered species or animals, living in zoos or tourism locations, can adopt invasive methods to detect emotions. The major advan- tages of adopting invasive methods are twofold: a) Invasive methods rely on more authentic measurements that relate to physiological indications of emotions; and, b) they de- liver accurate measurements while observing emotions in animals [ 51 ]. In non-invasive methods, emotions are collected using sensors mounted outside the bodies of animals. In fact, detecting emotions such as fear , anger , joy , disgust, neu- tral, and so forth, using facial expressions has been widely implemented for human datasets in the past [ 62 ]. Several methods and algorithms have been designed in the recent past to detect the emotions of human faces belonging to dis- persed geographical regions [ 43 ] [ 14 ]. Also, researchers Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 37 have captured the relationships between facial expressions and emotions [ 71 ]. As similar to identifying human emotions, classifying emotions from the facial expressions has been practiced by experienced animal researchers [ 29 ] [ 74 ]. They observed that adopting ICT technologies eases the process of emo- tion detection due to the inclusion of sophisticated learn- ing algorithms to capture the facial emotion-related features of animals. However , while investigating the existing re- search, we could observe that these mechanisms have tar - geted single animals rather than a group of animals while detecting their emotions. A lar ge sector of researchers worked on creating facial action coding systems (F ACS) in animals. Notably , F ACS was initially developed in human faces by observing fa- cial muscular movements [ 23 ]. Subsequently , the same ap- proach was developed in animals such as dogs, cats [ 12 ], and horse. [ 39 ] applied the F ACS approach to indicate pain points in animals; [ 73 ] evaluated the facial expressions of multiple animal species. In [ 19 ], authors attempted to label the facial features of animals before collecting their emo- tions. However , their approach led to reduced learning ac- curacy due to poor precision in labeling the F ACS of an- imals. In a few research works, researchers have applied deep learning algorithms to detect such facial expressions [ 40 ]. Facial expressions of animals are useful to capture and assess the severity of pain in animals. An array of re- search works is based on the grimace score of the facial expressions of animals. In short, the grimace score is an assessment score that indicates the pain levels of animals by observing their facial muscular movements. For in- stance, in [ 47 ], authors have utilized the grimace scale to detect the severity of emotions due to pain [ 46 ]. Simi- larly , the grimace scale indication has been applied in dif- ferent varieties of animals such as laboratory animals [ 67 ], farm animals [ 30 ], and pet animals [ 24 ] by a few other re- searchers to study the impact of pains in dif ferent animals – i.e., researchers [ 45 ] have extended the concept of gri- mace scales for sheep and named them as Sheep Pain Facial Expression Scale(SPFES) [ 45 ]; similarly , authors of [ 38 ] and [ 10 ] have developed unique grimace scales for horses and named them as Horse Grimace Scale (HGS). A few authors studied the relationship between animal behaviors and emotions. They monitored the behavior of multiple types of animals by assessing the behavior indica- tors [ 61 ]. For instance, the authors of [ 21 ] have developed a cow emotion estimation framework considering the va- lence of the af fective states of cow’ s behaviors; the authors of [ 33 ] have implemented a few behavior indicators of pigs based on the wagging of their tails; the authors of [ 20 ] have classified the emotions of horses after collecting their be- havior traits; authors have explored and af firmed the repet- itive pattern of horse movements that reduces their stress level [ 3 ]. Also, in [ 56 ], the authors have developed a mon- itoring framework that keeps track of the behavioral states of animals. Apart from farm or pet animals, we collected research works that focused on the behavioral study of elephants. Notably , researchers [ 66 ] and [ 8 ] have studied the behavior patterns of elephants using vocalization and vocal expres- sions. 2.4.1 Using pose detection Identifying the poses of animals and relating them to their core emotions is another aspect of detecting emotions in animals. The body posture and movements of animals con- vey emotional indices. The correlations between pose and emotions can be studied using modern learning algorithms, including deep learning algorithms. Such studies relating to animal emotions based on pose estimations can be instru- mental in minimizing pain in animals. For instance, in [ 77 ], the authors have developed a pose estimation system for dogs to assuage their pains. In the recent past, authors have developed a DeepLabCut frame- work [ 63 ] that relates the emotions of dogs, such as happi- ness, fear , and anger , to their corresponding poses. They have detected a few abnormalities in horses’ poses while considering their movement patterns. There are a few tools that estimate the poses of ani- mals and relate them to their emotional states. Notably , the DeepLabCut tool [ 44 , 26 ] and the LEAP tool [ 57 ] have been widely utilized among researchers and practitioners to capture the emotions of animals using deep learning algo- rithms. The DeepLabCut framework has also been utilized to generate emotion indicators from several cross-species of animals based on learned poses [ 50 ]. The summary of the existing works point out that most of the works have not applied specific cost-ef ficient solutions to detect animal emotions in real-time. T able 2.4 highlights the animal emotion detection systems that dif ferentiate the non-inclusion of cloud-based systems or IoT -enabled ap- proaches. In the table, we have definedN as NIL,NS as Not Scal- able,OFF as of fline,ON as online,S− Exp as standalone experimental setup,S− GPU as standalone experimental setup involving NVIDIA or similar GPU-based machines, Cloud as cloud-based solution, and S− Drone as stan- dalone experimental setup based on drones. The last metric denotes the cost ef ficiency in a scale ranking between 1 to 5 where 1 corresponds to the cost-ef ficient solution. The idea is to point out that serverless-based edge- enabled solution could be a better approach for detecting animal emotions that surpass a long time internal. 3 Edge-enabled serverless ar chitectur es for animal emotion detection W ith the alarming rise in interest in establishing human- animal bonds and preventing animal attacks, robust tech- nologies are required to detect the emotions of animals. 38 Informatica 49 (2025) 33–48 S. Benedict et al. T able 1: Animal emotion detection and allied techniques – a comparative study Article Methods and Pr ocedur es –Computing | Scalable | Offline/Online | Cost Scale [ 4 ] Manual analysis N | NS | OFF | 3 [ 51 ] YOLO and FasterRCNN for detecting cow emotions S-GPU | NS | OFF | 4 [ 1 ] Pre-trained CNN to study post sur gical impacts of mouse S-Exp | NS | OFF | 4 [ 10 ] Horse pain detection using machine learning S-Exp | NS | ON | 2 [ 17 ] Facial emotion detection of horses S-Exp | NS | OFF | 4 [ 26 ] Dog’ s emotion detection using neural network S-Exp | NS | OFF | 4 [ 29 ] Identify emotions in monkeys S-GPU | NS | OFF | 4 [ 33 ] T ail posture identification in pigs S-Drone | NS | ON | 4 [ 38 ] Pain in horses using CNN S-Exp | NS | OFF | 4 [ 58 ] Disease prediction using CNN S-Exp | NS | OFF | 4 [ 70 ] IoT -based elephant acoustic study using Neural network Cloud | S | ON | 2 [ 74 ] RetinaNet face posture identification S-GPU | NS | OFF | 4 [ 77 ] YOLO and LSTM-based pain detection in dogs S-Exp | NS | OFF | 4 [ 34 ] CNN-based animal face detection S-GPU | NS | OFF | 4 This section discusses the edge-enabled serverless architec- tures and frameworks that adopt an ef ficient animal emo- tion detection mechanism. T owards this end, at first, we cover the special features of architectures that are crucial to detecting animal emotions; next, we describe the possible software/hardware compo- nents and associated algorithms that improve the goals of identifying the emotions of animals; and, at last, a few met- rics that drive the objectives of architectures are discussed. 3.1 Animal emotion detection ar chitectur es In the past, specialized trainers detected animals’ emo- tions. They applied skilled tricks to improve the produc- tivity of the animals’ assigned tasks. For instance, vocal- izations such as growls, barks, or body postures such as wagging tails of dogs, are read by people raising animals. However , observing emotions and classifying the af fective states of animals using the manual approach can lead to hu- man bias [ 49 ]. There have been ef forts in the past to de- velop mechanisms that of fer unbiased assessments. How- ever , the works are under -researched to date [ 36 ]. This section elaborates on the applicability of IoT - enabled technologies for studying the emotions of animals. Depending on the deployment options, the animal emo- tion detection mechanism is classified into four types: i) Naive IoT , ii) Edge-enabled IoT , iii) Serverless IoT , and iv) AI-Assisted IoT (see Figure 2 ). Additionally , the sec- tion highlights the importance of involving deep learning- driven edge-enabled serverless architecture to enhance the accuracy and cost ef ficiencies. 3.1.1 Naive IoT – T ype-I Exploring the emotions of animals utilizes sensor -enabled networks that transport sensed data using communication protocols such as W iFi, 4G/5G, or Long Range W AN (Lo- RA W AN). The W iFi communication protocol provides a higher bandwidth to transfer sensor data from a camera. This protocol is suitable in locations such as zoos, tem- ples, and work sites, where the emotions of animals need to be monitored. 4G/5G networks are mobile cellular net- works based on 4G/5G communication protocols that of- fer a high-speed network to carry animal images or video streams in a wireless medium. However , this protocol is not suitable for carrying sensory information for long dis- tances. On the contrary , LoRA W AN is a wide area network that can transfer sensory information that reaches around 10 KMs between sensors and services. The drawback of utilizing the LoRaW AN protocol in emotion detection ar - chitectures is its limitation in transferring a huge volume of data to sense the emotions of animals. In the Naive-IoT system, sensors such as camera sensors collect animal images or frames and stream data to cloud services through gateways that are connected using com- munication protocols such as W iFi or 4G/5G networks. The cloud services process these animal videos in cloud envi- ronments after suf ficient data processing mechanisms, in- cluding data filtering, augmentation, and so forth. Addi- tionally , the cloud services host learning algorithms or AI- assisted services to detect the emotions of animals, such as anger , grief, happiness, joy , and so forth. For instance, in [ 54 ], the authors have developed an IoT - enabled electronic board using sensors that were mounted on the ears of pigs to observe their activities; [ 41 ] have de- veloped an accelerator -based solution to collect the behav- ioral changes of horses by designing a few af fective states; [ 69 ] have studied the emotions of animals by transforming learning algorithms on smartwatches; and so forth. Simi- larly , [ 52 ] have surveyed the sensor -based emotion moni- toring approaches for farm animals, which limit the scope of research to a subset of animals. In Figure 2 , we could ob- serve that the T ype-I architecture has direct communication of sensors to cloud environments. Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 39 Figure 2: Animal emotion detection architectures – T ype-I, II, III, and IV 3.1.2 Edge-enabled IoT – T ype-II Rather than streaming animal videos from their exact loca- tions, such as the zoo or circus arena to the cloud environ- ments, it is better to transfer the important features of ani- mals’ emotions to edge nodes. This enables us to minimize the network bandwidth and privacy of streaming sites. For instance, to observe the emotions of elephants, while learn- ing their emotions, it is suf ficient to submit features such as ear positions, mouth angle, and so forth, to cloud services based on animal emotion detection services. Most of the video processing and parsing of facial features or emotional traits could be carried out on edge devices that are mounted by decent processing elements. The most notable edge de- vices that are utilized for processing video frames nearer to the sensor nodes are Raspberry Pi nodes, mobile phones, low computational servers, laptops, Jetson Nano devices, and so forth. The type-II architecture model of Figure 2 rep- resents the sensors connected to clouds through edge nodes. A few researchers have developed frameworks with multi- sensor devices, including cameras, to detect the emotions of elephants on smartphone devices based on GPS communi- cations [ 70 ]. However , this work is not assisted by modern communication protocols or edge computing services to ef- ficiently utilize the power of edge-enabled IoT systems. In fact, edge computing nodes suf fer from memory and computation-related resource limitations. They could not handle memory-expensive deep learning tasks such as ResNet50. 3.1.3 Serverless IoT – T ype-III Most IoT -enabled applications or pure edge-enabled so- lutions continuously power the associated cloud services. This approach of utilizing edge and cloud computing nodes for processing animals’ emotions has three major disadvan- tages: 1. Ener gy Inefficiency – Cloud services are unnecessar - ily underutilized, even if there is no data to process them. For instance, the videos of animals are not cap- tured throughout the day/year to study the emotional variations over time in the zoo. The emotions could be programmatically observed only for a few minutes, especially while feeding, playing, or so forth. In such situations, cloud servers could be switched “OFF“ to ef ficiently handle the ener gy consumption of these ap- plications. Also, in locations where renewable ener gy sources are available, it is suf ficient to utilize renew- able ener gy sources to power “ON“ the devices. 2. Cost Inefficiency – In cloud environments, compute instances and their allied resources for applications are continually utilized. If we need to detect the pat- terns of emotions in animals over months or years in a zoo/circus, the cloud resources are always utilized, although the necessity of resources is limited. This in- directly increases the costs of such emotion detection- based applications for users. 3. Carbon Inefficiency – When animal emotion-related applications are executed on clouds, they heat the cloud resources. Obviously , ener gy consumption due to powering “ON“ the machines and cooling them due to overutilization the risk of carbon emissions in data centers or supercomputers. These carbon emissions, therefore, could be controlled for the benefit of soci- ety if the applications are programmed elegantly . 40 Informatica 49 (2025) 33–48 S. Benedict et al. T able 2: Deep learning algorithm-based animal emotion de- tection Reference Animal Algorithm Approach Location [ 59 ] Horses Object Detection NN and YOLO Facial Keypoint Action Movement Detection Horse Staple [ 2 ] Horses LSTM concatenated binary classifier Facial Keypoint Action Movement Detection Horse Staple [ 51 ] Cows, Pigs FasterRCNN YOLOv3 YOLOv4 Emotion Classification using dif ferent Algorithms Pasture [ 27 ] Dogs Pretrained AlexNet with FC Layers Emotion Pretraining - [ 26 ] Dogs DeepLabCut Landmark points and posture detection (Ear , Hair , Mouth, and so forth) - [ 1 ] Mice ResNet50 Emotions post sur gical Laboratory [ 17 ] Horses FasterRCNN VGG16 ResNet50v2 Xception Comparison of algorithms based on head positions Farm land [ 7 ] Dogs ResNet Supervised Classify emotions based on anticipation or frustration Home [ 58 ] Sheep YOLO SSDMobileNet Sheep face detection Farm land [ 38 ] Horses 3-layered CNN Facial feature classification based on Grimmascale Farm land [ 10 ] Horses InceptionV3 VGG+LSTM Pain detection with temporal features Farm land [ 60 ] Horses Encoder - Decoder Equine pain classification Farm land A serverless-enabled cloud implementation model im- proves ener gy ef ficiency , cost ef ficiency , and carbon ef fi- ciency . It is a cloud execution model where servers are not powered “ON“ throughout the execution of applications. Only if sensor data is triggered using edge-enabled devices, the servers are powered “ON“ and the states of the applica- tions are modified. For instance, the emotions of animals are assessed based on the qualified input data captured and transferred to cloud servers. There exist a few platforms that enable serverless- assisted IoT solutions. For instance, A WS of fers Lambda services. The Lambda services enable execution of func- tions based on triggers obtained from IoT -enabled sensors. 3.1.4 AI-assisted IoT solutions – T ype-IV Providing AI-based solutions is crucial to accurately detect- ing the emotions of animals. AI-based algorithms, in gen- eral, are classified into supervised, unsupervised, and rein- forcement learning algorithms [ 6 ]. They are applied in var - ious domains, such as the financial sector , agricultural sec- tor , cognitive sector [ 15 ], and so forth. Supervised learning algorithms are further classified based on machine learning approaches and deep learning approaches. There are several classical machine learning algorithms such as Support V ector Machines [ 39 ], Random Forests, K-Nearest Neighbor , Rule-based, Naive Bayes, Principal Component Analysis, and so forth, that could be adopted for identifying the emotions of animals from varied input sources such as texts, video, or audio sources. Notably , authors have utilized the vocalization features of elephants [ 70 ] to detect the state of mind using SVM classifiers in four dif ferent categories: roar , rumble, trumpet, and cry . Apart from a few classical ML algorithms, there exist a few deep learning algorithms such as the Y ou Only Look Once (YOLO) algorithm, Convolutional Neural Networks (CNNs), Residual Net (ResNET) models, Reinforcement Neural Networks (RNN), and so forth for detecting animal emotions. For instance, in the recent past, the authors of [ 31 ] have assessed the pain in horses [ 2 ] and donkeys by detecting the facial keypoints; the authors of [ 25 ] have applied the ResNet50 algorithm to detect pain in cats; researchers of [ 45 ] have utilized computer vision approaches to detect pain in animals; in [ 28 ], the authors have implemented an IoT -enabled solution to establish an intelligent ecosystem for tracking the health of diary farm animals in real time. Similarly , authors of [ 72 ] have implemented a Faster R-CNN deep learning algorithm to classify animal emo- tions; a few authors have observed the emotions of farm animals using pre-trained Deep CNN models when pro- jected with wolf sound and images [ 51 ]; a few researchers, [ 27 ], have developed a simple pre-trained AlexNet deep learning model to classify the af fective states, namely , smiling, growling, and sleeping states of dogs; a few other researchers [ 1 ] have developed an emotion detection model based on neural networks supported by binary classi- fiers that detected mice’ s post-anesthetic sur gical emotions based on facial expressions; a few researchers [ 17 ] have studied the facial expressions of horses using CNNs; a few authors have worked on the application of deep learning al- gorithms for improving the emotion intelligence of dogs [ 7 ] and cats [ 25 ]; and, so forth. Additionally , authors have utilized a pre-trained ResNet- based deep learning model to classify the emotions of dogs as anticipation or frustration [ 7 ]; authors [ 58 ] have developed temporal features as input to detect the emo- tions of sheep using multi-step CNN models; in [ 38 ], re- searchers have implemented 3-layered CNN-based deep learning models to recognize pain in horses using gri- mace scales; similarly , several researchers have applied temporal features in deep learning algorithms, including encoder -decoder -based self supervised deep learning algo- rithms [ 10 ] [ 1 1 ] [ 60 ], to detect the emotions of animals. Based on the most commonly utilized deep learning algo- rithms, we classified and tabulated them in T able 2 . Obser - vations reveal a lack of extensive research on the resource or cost ef ficiency of learning methods in animal emotion detection. Interestingly , by integrating these potential deep learn- ing algorithms into edge-enabled serverless frameworks, we can quickly , accurately , and ef ficiently detect animal Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 41 emotions when sensors such as cameras trigger sensor data in real-time. 3.2 Ar chitectur e evaluation metrics The animal emotion detection architectures such as Naive IoT , Edge-enabled IoT , Serverless IoT , or AI-Assisted IoT have to be monitored and evaluated. W e identified a set of performance metrics that could be utilized in the architec- tures. The description of these metrics is explained below: Device utility The device utility is a measure of the to- tal utilization of cloud or sensor -based resources among the available utilization times. In serverless environments, the resources are utilized only when there is a trigger from asso- ciated sensors – i.e., the utilization depends on the nature of applications, serverless platforms, and resource providers’ serviceability . Configurational easiness Configuring a serverless- enabled IoT platform can lead to dif ficulties for some users. For instance, a few serverless platforms utilize sophisticated user interfaces or ease in setting up automatic configurations, whereas, others do not. Hence, depending on the ease of setting up the configurations, we classify the architectures. Multi-communication support An animal emotion de- tection architecture is limited to supporting all communi- cation protocols. Some architectures enable various com- munication protocols such as, W iFi, Bluetooth, 4G/5G, and so forth. However , the cost of the system increases if these devices operate with dif ferent communication pro- tocols. For ease, based on the utility of the application, several IoT -enabled emotion detection architectures are de- signed using W iFi, 4G/5G, or Bluetooth. This metric en- sures the ability of the device to install it on any dif ferent locations/platforms. Pr ediction accuracy The emotion detection of elephants or animals in farms or in the wild involves learning algo- rithms, including deep learning algorithms such as CNNs, LSTM, or SVM. While classifying emotions depending on the environmental inputs, prediction accuracy is one cru- cial metric that needs to be maximized. If not, false pos- itives drive the system with wrong decisions. Developing an emotion detection system that of fers more accurate deci- sions is a time-consuming task, as it depends on the quality of the learning algorithms and the learning parameters of the algorithms. Energy or hardwar e-r elated metrics The performance of applications is a step toward improving the implementa- tion strategies and the writing style of applications. In an edge-enabled elephant emotion detection system, develop- ers could implement algorithms, including learning algo- rithms, in several approaches: 1. Lightweight Implementation using Containers – container -based implementations of fer lightweight solutions that self-contain all relevant packages to ex- ecute emotion detection applications. It is an OS-level virtualization approach in which algorithmic instances are of floaded to multiple servers depending on the load balancing requirements or other performance concerns of the execution environments. 2. Security-conscious Implementation Appr oach – Ap- plications could be developed considering the secu- rity features of the programming models. Increasing the security-based software components in emotion- related detection algorithms can reduce the perfor - mance of applications. 3. Execution T ime Impr ovements – Improving the ex- ecution time of learning algorithms, including deep learning-based emotion detection algorithms, requires the application developer to have algorithmic excel- lence and coding skillsets. Developing algorithms considering the nature of input data and avoiding un- necessary allocation of variables could improve the execution time of emotion detection applications. 4. Ener gy or Carbon-conscious Developments – Met- rics that improve the ener gy ef ficiency of applica- tions or carbon emissions are predominantly prac- ticed in long-running applications. For instance, edge-enabled emotion detection applications involve battery-operated or renewable ener gy sources for pow- ering edge or sensor nodes. The entire ecosystem should consciously deliver most of the executions to such renewable-powered resources while learning emotions or inferencing emotion-related patterns. In this way , carbon emissions are reduced in such long- running applications, especially when training phases are carried out on edge-enabled devices. Figure 2 illustrates the dif ferent types of animal emo- tion detection architectures depending on the deployment options – i.e., whether they involve edge-based solutions, serverless execution models, or naive IoT -enabled ap- proaches. 3.3 DL-driven edge-enabled serverless ar chitectur e – an exploratory study Among the four types of animal emotion detection architec- tures, the T ype-IV architecture that combines deep learn- ing, IoT , edge computing, and serverless execution models is considered to be more ef ficient in terms of cost and re- source utilization. 42 Informatica 49 (2025) 33–48 S. Benedict et al. T able 3: Cost exploration for A WS services in naive-IoT architectures A WS Services Name of the Service Dollars per hour Y ear -wise Compute Instances t4g.micro 0.0104 91.104 m5.lar ge 0.096 840.96 t2.micro 0.01 16 101.616 g4dn.12xlar ge instances (4 GPUs - 96vCPUs) 5.47 47917.2 Associated Storage SSD-32GB 2.56 30.72 IOPS-SSD-32GB 4 48 Kinesis Data fir ehose 1TB-streams 29.7 356.4 A WS S3 100GB storage 2.3 27.6 T otal (10 instances) 12556.8 T able 4: Comparison of types of animal emotion detection architectures Ar chitectur e Scalable Latency Cost Efficiency Security Accuracy Utility Easiness Manual Poor Depends Poor High Depends NA NA on availability on Evaluator T ype-I Medium High Poor Poor Depends Poor Poor on Application T ype-II High Medium Medium High Depends Medium Poor on Application T ype-III High Low High High Depends High High on Application T ype-IV High Low High High Depends on AI High Medium 3.3.1 Pr ocess involved The following steps outline the process of incorporat- ing deep learning algorithms into edge-enabled serverless frameworks within the Amazon A WS ecosystem: 1. Initially , camera sensors stream videos to nearby edge nodes for evaluating the animals interested in detect- ing the emotions. For instance, an elephant emotion detection system captures frames belonging to ele- phants and omits the other animals. 2. Next, edge-enabled services connect to A WS S3 buck- ets to trigger the required DL algorithm, such as YOLO, CNN, RESNET , or so forth, using S3 event notification options. The S3 buckets include code snippets and scripts to start executing EC2 instances based on A WS Lambda functions. 3. Accordingly , the corresponding DL algorithm is exe- cuted along with the past state information and data modules on EC2 compute instances for a specified time interval. The EC2 service requests are activated through simple queue services of A WS. 4. Finally , once the modeling and prediction tasks are handled, the state of the application is saved in the S3 bucket for future executions. It could be noticed that the EC2 instances remain inactive throughout the an- imal emotion detection processes, particularly when they are installed in forest or zoo locations. This dras- tically reduces the computational costs involved in the execution. Additionally , a few lightweight deep learn- ing algorithms, such as T inyYOLO, could be deployed for faster and more resource-ef ficient learning or infer - ences. 3.3.2 Cost-efficiency T o illustrate the cost ef ficiency of T ype-IV architecture – specifically , DL-driven edge-enabled serverless architec- ture, we have considered an A WS platform consisting of 10 Elastic Compute Cloud (EC2) instances, including GPU instances, to execute parallel deep learning models; A WS Lambda and A WS S3 services have been considered to es- tablish serverless implementations. Also, we have assumed that the architecture included cameras and edge devices such as Raspberry Pi nodes that are dedicatedly available for evaluating the costs involved in the animal emotion de- tection. In the study , only recurring costs involved in the operations were studied, assuming the animal emotion de- tection has to be evaluated throughout an year . The key findings are listed below: 1. The manual approach to detecting animal emotions is not a scalable or accurate solution as it is dependent on the skillset of evaluators; 2. The T ype-I architecture utilized cameras to stream animal videos directly to cloud-based services. In this approach, at least cloud instances have to be active throughout the execution of services. As an exploratory study , we considered ten m5.large in- stances from the US East (N. V ir ginia) region to eval- uate the animal emotions. Accordingly , the cost of the EC2 instances was estimated – i.e., it reached around $101.616 per year . Additionally , the architecture re- quired high performance storage units to quickly re- Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 43 spond to the requests for animal emotion detections. Hence, Solid State Drives (SSDs) or I/O Operations Per Second (IOPS) storage units have to be invoked. Also, we need to process streaming data on the cloud. T o do so, A WS Kinesis Data Firehose have to be plugged in to the detection system; 3. Involving ten m5.large instances, 32-GB SSD, 1TB- streaming support, and 100GB S3 service can lead to $12556.8 for a year . T able 3 lists the required cloud services and the recurring cost required per year to op- erate an animal emotion detection-related application on the cloud using T ype-I architecture. 4. The T ype-II architecture considered edge-enabled de- vices to parse videos and perform data-related opera- tions closer to the data sources. This architecture still requires cloud-based compute instances and storage units mentioned in T able 3 to evaluate animal emo- tions. Although a few computations happen on the edge, the cloud services have to be switched ‘ON‘ throughout the year . However , the network bandwidth of the application is relatively improved when com- pared to T ype-I architecture. 5. T ype-III and T ype-IV architectures involve serverless IoT solutions. Here, the A WS Lambda service is integrated to the animal emotion detection applica- tion. In doing so, EC2 compute instances are uti- lized only based on the trigger happening from data sources. Assuming that there are 10,000 service re- quests prompted from sensor sites, there are no char ges collected specifically from A WS Lambda services. However , the T ype-III architecture requires EC2 in- stances to update weights and perform inferences. If fifteen minutes are required for performing each ser - vice request, the recurring costs of T ype-III architec- ture for an entire year will reach only $2676 – i.e., 10,000 times of ten m5.large instances along with 100GB storage components. The major dif ference be- tween T ype-III and T ype-IV architectures is the in- clusion of core deep learning models to perform ani- mal emotion detections. Accordingly , the accuracy of T ype-IV architecture increases with almost equivalent costs found in T ype-III architectures. 6. Additionally , the four types of architectures are com- pared in terms of scalability , latency , security , device utility , and ease. T able 4 highlights the importance of AI-assisted architecture using serverless and IoT com- ponents to detect animal emotions. 4 Resear ch dir ections This section suggests a few research directions that could be adopted based on the existing animal emotion detection strategies. The possible research directions are discussed in three major divisions, as shown in Figure 3 . Figure 3: Research Directions in Edge-Enabled Animal Emotion Domain 4.0.1 Impr oving learning accuracy Emotion detection using edge-enabled serverless architec- tures involves learning algorithms such as CNNs or YOLO. The learning accuracy of these algorithms could be in- creased by applying a few methods, as listed below: 1. the developers could design novel deep learning algo- rithms that increase the learning accuracy of applica- tions; 2. an automatic tuning of learning hyper -parameters such as learning rate, number of iterations, model optimiza- tion methods, and so forth, could be developed. In fact, developing an automatic tuning feature can lead to poor performance ef ficiencies while detecting ani- mal emotions; 3. an approach to compressing learning models while transferring them to the cloud or edge increases the network bandwidth of the learning system. Unfortu- nately , researchers avoid them as networks are consid- erably working well during their experiments with a low volume of data. However , the real situation man- dates the necessity of such compressions. 4.0.2 Computational advancements Although we proposed edge-enabled serverless-oriented architectures for animal emotion detection, there are re- search directions to improve them i) by choosing where 44 Informatica 49 (2025) 33–48 S. Benedict et al. to place the services, such as learning-related services or security-oriented services; ii) by designing tightly coupled hardware-software designs tar geting performance metrics such as ener gy , execution time, or carbon emissions; iii) by developing hierarchical architectures that benefit from the most commonly observed time and space complexities in computing environments; and, iv) by implementing edge- enabled serverless architectures that improve the cost ef fi- ciency factors and latency issues of animal emotion detec- tion algorithms. 4.0.3 Applications Apart from developing technologies and solutions to en- hance research in computational and algorithmic design, there are immense research opportunities to observe emo- tions among a swarm of animals. For instance, identifying animals’ emotions from geographically dispersed locations is a challenge. This involves an architectural design of the system, considering sensors and associated communication protocols to transfer intelligence elegantly . Similarly , in- ferring intelligence based on a group of multi-species ani- mals could spot novel findings. 5 Conclusions Animal emotion detection research has started with the fresh challenge of detecting emotions. T raditionally , a manual approach to detecting animal emotions had fail- ures and inef ficiencies, as pointed out by several prac- titioners. The recent IoT -enabled animal emotion detec- tion systems have marked an array of performance im- provement opportunities. This article explored the recent trends in applying IoT -enabled solutions for animal emo- tion detection research; it studied the importance of incor - porating DL-driven edge-enabled serverless-oriented archi- tectures by exploring the costs involved in computations of architectures; and, it of fered a few research directions that threw light on near -future research works. The article would enable the incorporation of novel architectural de- signs and significantly increase the interest of indigenous animal emotion researchers. Acknowledgement The authors thank IIIT -Kottayam for of fering resources, Prof. Michael Gerndt of TUM for motivating the work, and the editorial of fice the journal for reviewing/publishing the work. Refer ences [1] Andresen, N., Wöllhaf, M., Hohlbaum, K., Lewejo- hann, L., Hellwich, O., Thöne-Reineke, C., & Belik, V . (2020). T owards a fully automated surveillance of well-being status in laboratory mice using deep learn- ing: Starting with facial expression analysis. PloS one , (V ol. 15(4), e0228059, pp. 1-23). https://doi. org/10.1371/journal.pone.0228059 [2] Andersen, P . H., Broomé, S., Rashid, M., Lundblad, J., Ask, K., Li, Z., Hernlund, E., et al. (2021). T o- wards Machine Recognition of Facial Expressions of Pain in Horses. Animals , (V ol. 1 1, No.6, 1643, pp. 1-18). MDPI AG . Retrieved from http://dx.doi. org/10.3390/ani11061643 [3] Auer U, Kelemen Z, Engl V , Jenner F (2021) Ac- tivity T ime Budgets—A Potential T ool to Monitor Equine W elfare? Animals (V ol. 1 1, No. 850, pp. 1- 12). https://doi.org/10.3390/ani11030850 [4] Battini M., Agostini A., & Mattiello, S. (2019). Un- derstanding Cows’ Emotions on Farm: Are Eye White and Ear Posture Reliable Indicators?, Animals , (V ol. 9, No. 8, 477, pp. 1-12). https://doi.org/10.3390/ ani9080477 [5] Bekof f, M. (2023). Human Emotions in An- imals, https://online.uwa.edu/news/ empathy- in- animals/ , Accessed in Oct. 2024. [6] Benedict, S. (2022). Deep Learning T echnologies for Social Impact. IOP Publishing . pp. 1-13, https:// doi.org/10.1088/978- 0- 7503- 4024- 3 [7] Boneh-Shitrit, T ., Amir , S., Bremhorst, A., Mills, D., Riemer , S., Fried, D., & Zamansky , A. (2022). Deep Learning Models for Automated Classification of Dog Emotional States from Facial Expressions. in arXiv:2206.05619 , pp. 1-6, https://doi.org/10. 48550/arXiv.2206.05619 [8] Briefer , E. F . (2012). V ocal expression of emotions in mammals: mechanisms of production and evidence. In S. Le Comber (Ed.), Journal of Zoology (V ol. 288, Issue 1, pp. 1–20). W iley Portico . https://doi. org/10.1111/j.1469- 7998.2012.00920.x [9] Bremhorst, A., Sutter , N. A., Würbel, H., Mills, D. S., & Riemer , S. (2019). Dif ferences in facial expressions during positive anticipation and frustra- tion in dogs awaiting a reward. In Scientific Reports . Springer Science and Business Media LLC. (V ol. 9, No. 19312, pp. 1-13). https://doi.org/10.1038/ s41598- 019- 55714- 6 [10] Broomé, S., Gleerup, K. B., Andersen, P . H., & Kjellstrom, H. (2019). Dynamics Are Important for the Recognition of Equine Pain in V ideo. In 2019 IEEE/CVF Confer ence on Computer V ision and Pattern Recognition (CVPR) . IEEE. USA. pp. 12659-12668. https://doi.org/10.1109/cvpr. 2019.01295 Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 45 [1 1] Broomé, S., Ask, K., Rashid-Engström, M., Haubro Andersen, P ., & Kjellström, H. (2022). Sharing pain: Using pain domain transfer for video recognition of low grade orthopedic pain in horses. In A. Seal (Ed.), PLOS ONE (V ol. 17, No. 3, e0263854, pp. 1-21). Pub- lic Library of Science (PLoS). https://doi.org/ 10.1371/journal.pone.0263854 [12] Caeiro, C. C., Burrows, A. M., & W aller , B. M. (2017). Development and application of CatF ACS: Are human cat adopters influenced by cat facial ex- pressions? In Applied Animal Behaviour Science . (V ol. 189, pp. 66–78). Elsevier BV . https://doi. org/10.1016/j.applanim.2017.01.005 [13] Camerlink, I., Coulange, E., Farish, M., Baxter , E. M., & T urner , S. P . (2018). Facial expression as a poten- tial measure of both intent and emotion. In Scientific Reports (V ol. 8, Issue 17602). Springer Science and Business Media LLC. pp. 1-9. https://doi.org/ 10.1038/s41598- 018- 35905- 3 [14] Chen, C., Zhu, W ., Steibel, J., Siegford, J., W urtz, K., Han, J., & Norton, T . (2020). Recognition of ag- gressive episodes of pigs based on convolutional neu- ral network and long short-term memory . In Com- puters and Electr onics in Agricultur e . (V ol. 169, No. 105166, pp. 1-10). Elsevier BV . pp. 1-10. https: //doi.org/10.1016/j.compag.2019.105166 [15] Cheng Pan, Haiyan Zhao, Meijiao Sun. (2024). Real- time T ar get Detection System in Scenic Landscape Based on Improved YOLOv4 Algorithm. in Infor - matica (V ol. 48, p. 35–48). https://doi.org/10. 31449/inf.v48i8.5700 [16] Clemins, P . J., & Johnson, M. T . (2003). Applica- tion of speech recognition to African elephant (Lox- odonta africana) vocalizations. Pr oceedings. (ICASSP ’03). International Confer ence on Acoustics, Speech and Signal Pr ocessing (ICASSP’03) . IEEE. pp. I-484- 487. https://doi.org/10.1109/icassp.2003. 1198823 [17] Corujo, L. A., Kieson, E., Schloesser , T ., & Gloor , P . A. (2021). Emotion Recognition in Horses with Con- volutional Neural Networks. In Futur e Internet . (V ol. 13, Issue 10, p. 250). MDPI AG. https://doi.org/ 10.3390/fi13100250 [18] Cowen AS, Keltner D. (2017). Self-report captures 27 distinct categories of emotion bridged by continuous gradients. Pr oc Natl Acad Sci U S A . (V ol. 1 14, No. 38, p. E7900-E7909. https://doi.org/10.1073/ pnas.1702247114 [19] Dai, F ., Leach, M., MacRae, A. M., Minero, M., & Dalla Costa, E. (2020). Does Thirty-Minute Standard- ised T raining Improve the Inter -Observer Reliability of the Horse Grimace Scale (HGS)? A Case Study . In Animals (V ol. 10, No. 5, 781, pp. 1-7). MDPI AG. https://doi.org/10.3390/ani10050781 [20] Dyson, S., & Pollard, D. (2020). Application of a Rid- den Horse Pain Ethogram and Its Relationship with Gait in a Convenience Sample of 60 Riding Horses. In Animals (V ol. 10, No. 6, 1044, pp. 1-14). MDPI AG. https://doi.org/10.3390/ani10061044 [21] Ede, T ., Lecorps, B., von Keyserlingk, M. A. G., & W eary , D. M. (2019). Symposium review: Sci- entific assessment of af fective states in dairy cattle. In Journal of Dairy Science (V ol. 102, No. 1 1, pp. 10677–10694). American Dairy Science Association. https://doi.org/10.3168/jds.2019- 16325 [22] Ekman, P . (2023). Basic Emotions. https: //www.paulekman.com/wp- content/uploads/ 2013/07/Basic- Emotions.pdf , Accessed Jan. 2023. [23] Ekman, P ., & Friesen, W . V . (1978). Facial Ac- tion Coding System. In PsycTESTS Dataset . Amer - ican Psychological Association (AP A). https:// doi.org/10.1037/t27734- 000 [24] Evangelista, M. C., W atanabe, R., Leung, V . S. Y ., Monteiro, B. P ., O’T oole, E., Pang, D. S. J., & Steagall, P . V . (2019). Facial expressions of pain in cats: the development and validation of a Feline Grimace Scale. In Scientific Reports (V ol. 9, No. 19128, pp.1-1 1). Springer Science and Business Media LLC. https://doi.org/10. 1038/s41598- 019- 55693- 8 [25] Feighelstein, M., Shimshoni, I., Finka, L. R., Luna, S. P . L., Mills, D. S., & Zamansky , A. (2022). Au- tomated recognition of pain in cats. In Scientific Re- ports (V ol. 12, No. 9575, pp.1-10). Springer Science and Business Media LLC. https://doi.org/10. 1038/s41598- 022- 13348- 1 [26] Ferres, K., Schloesser , T ., & Gloor , P . A. (2022). Predicting Dog Emotions Based on Posture Analy- sis Using DeepLabCut. In Futur e Internet (V ol. 14 (4) No. 97, p. 1-16). MDPI AG. https://doi.org/10. 3390/fi14040097 [27] Franzoni, V ., Milani, A., Biondi, G., & Micheli, F . (2019). A Preliminary W ork on Dog Emotion Recog- nition. In IEEE/WIC/ACM International Confer ence on W eb Intelligence - Companion V olume. WI ’19: IEEE/WIC/ACM International Confer ence on W eb In- telligence . ACM. pp. 91-96, https://doi.org/10. 1145/3358695.3361750 [28] Gehlot, A., Malik, P . K., Singh, R., Akram, S. V ., & Alsuwian, T . (2022). Dairy 4.0: Intelligent Com- munication Ecosystem for the Cattle Animal W elfare with Blockchain and IoT Enabled T echnologies. In 46 Informatica 49 (2025) 33–48 S. Benedict et al. Applied Sciences (V ol. 12, 7316, p.1-19). MDPI AG. https://doi.org/10.3390/app12147316 [29] Guo, S., Xu, P ., Miao, Q., Shao, G., Chapman, C. A., Chen, X., He, G., Fang, D., Zhang, H., Sun, Y ., Shi, Z., & Li, B. (2020). Automatic Identifica- tion of Individual Primates with Deep Learning T ech- niques. In iScience (V ol. 23, No.8, 101412, p.1-9). El- sevier BV . https://doi.org/10.1016/j.isci. 2020.101412 [30] Häger , C., Biernot, S., Buettner , M., Glage, S., Keubler , L. M., Held, N., Bleich, E. M., Otto, K., Müller , C. W ., Decker , S., T albot, S. R., & Ble- ich, A. (2017). The Sheep Grimace Scale as an in- dicator of post-operative distress and pain in labo- ratory sheep. In I. A. S. Olsson (Ed.), PLOS ONE (V ol. 12, Issue 4, e0175839, p.1-15). Public Library of Science (PLoS). https://doi.org/10.1371/ journal.pone.0175839 [31] Hummel, H. I., Pessanha, F ., Salah, A. A., van Loon, T . J. P . A. M., & V eltkamp, R. C. (2020). Auto- matic Pain Detection on Horse and Donkey Faces. In 2020 15th IEEE International Confer ence on Au- tomatic Face and Gestur e Recognition (FG 2020). 2020 . IEEE. pp. 793-800, https://doi.org/10. 1109/fg47880.2020.00114 [32] ICARUS, (2023). ICARUS Global Mon- itoring. Animals Early W arning System. https://www.icarus.mpg.de/28810/ animals- warning- sensors , accessed in Oct. 2024. [33] Iglesias, P . M., & Camerlink, I. (2022). T ail posture and motion in relation to natural behaviour in juvenile and adult pigs. In Animals (V ol. 16, Issue 4, 100489, p. 1-10). Elsevier BV . https://doi.org/10.1016/ j.animal.2022.100489 [34] Islem Jarraya, Fatma BenSaid, W ael Ouarda, Umapada Pal and Adel M. Alimi. (2023). A New Convolutional Neural Network based on a Sparse Convolutional Layer for Ani- mal Face Detection. In Multimedia T ools and Applications (V ol. 82, p. 91-124). https: //doi.org/10.1007/s11042- 022- 12610- y [35] João V ideira, Pedro D. Gaspar , V asco N. G. J. Soares, João M. L. P . Caldeira. (2024). A Mobile Applica- tion for Detecting and Monitoring the Development Stages of W ild Flowers and Plants. In Informatica . (V ol 48, No 6, pp. 43-58), https://doi.org/10. 31449/inf.v48i6.5645 [36] Kret, M. E., Massen, J. J. M., & de W aal, F . B. M. (2022). My Fear Is Not, and Never W ill Be, Y our Fear: On Emotions and Feelings in Animals. In Affective Science (V ol. 3, Issue 1, pp. 182–189). Springer Sci- ence and Business Media LLC. https://doi.org/ 10.1007/s42761- 021- 00099- x [37] Langbauer , W . R. (2000). Elephant communication. In Zoo Biology (V ol. 19, No. 5, pp. 425–445). W iley . https://doi.org/10.1002/1098- 2361(2000) 19:5< 425::AID- ZOO11> 3.0.CO;2- A [38] Lencioni, G. C., de Sousa, R. V ., de Souza Sardinha, E. J., Corrêa, R. R., & Zanella, A. J. (2021). Pain assessment in horses using automatic facial expression recognition through deep learning-based modeling. In H . Nisar (Ed.), PLOS ONE (V ol. 16, No. 10, e0258672, p. 1-12). Public Library of Science (PLoS). https://doi.org/10.1371/ journal.pone.0258672 [39] Lundblad, J., Rashid, M., Rhodin, M., & Haubro An- dersen, P . (2021). Ef fect of transportation and social isolation on facial expressions of healthy horses. In E. Palagi (Ed.), PLOS ONE (V ol. 16, No. 6, e0241532, p. 1-17). Public Library of Science (PLoS). https: //doi.org/10.1371/journal.pone.0241532 [40] Li, Z., Broomé, S., Andersen, P .H., & Kjellström, H. (2021). Automated Detection of Equine Facial Ac- tion Units. ArXiv , p. 1-5. https://arxiv.org/abs/ 2102.08983 [41] Maisonpierre, I. N., Sutton, M. A., Harris, P ., Men- zies￿Gow , N., W eller , R., & Pfau, T . (2019). Ac- celerometer activity tracking in horses and the ef fect of pasture management on time budget. In Equine V et- erinary Journal (V ol. 51, No. 6, pp. 840–845). W iley . https://doi.org/10.1111/evj.13130 [42] Mendl, M., Neville, V ., & Paul, E. S. (2022). Bridg- ing the Gap: Human Emotions and Animal Emotions. In Affective Science (V ol. 3, No. 4, pp. 703–712). Springer Science and Business Media LLC. https: //doi.org/10.1007/s42761- 022- 00125- 6 [43] Marsot, M., Mei, J., Shan, X., Y e, L., Feng, P ., Y an, X., Li, C., & Zhao, Y . (2020). An adaptive pig face recognition approach using Convolutional Neural Networks. In Computers and Electr onics in Agricultur e (V ol. 173, No. 105386, p. 1-10). Else- vier BV . https://doi.org/10.1016/j.compag. 2020.105386 [44] Mathis, A., Mamidanna, P ., Cury , K. M., Abe, T ., Murthy , V . N., Mathis, M. W ., & Bethge, M. (2018). DeepLabCut: markerless pose estimation of user - defined body parts with deep learning. In Natur e Neu- r oscience (V ol. 21, No. 9, p. 1281–1289). Springer Science and Business Media LLC. https://doi. org/10.1038/s41593- 018- 0209- y Deep Learning-Driven Edge-Enabled Serverless… Informatica 49 (2025) 33–48 47 [45] McLennan, K., & Mahmoud, M. (2019). Develop- ment of an Automated Pain Facial Expression Detec- tion System for Sheep (Ovis Aries). In Animals (V ol. 9, No. 4, 196, p. 1-7). MDPI AG. https://doi. org/10.3390/ani9040196 [46] Mogil, J. S., Pang, D. S. J., Silva Dutra, G. G., & Chambers, C. T . (2020). The development and use of facial grimace scales for pain measurement in an- imals. In Neur oscience, Biobehavioral Reviews (V ol. 1 16, pp. 480–493). Elsevier BV . https://doi.org/ 10.1016/j.neubiorev.2020.07.013 [47] Mota-Rojas, D., Olmos-Hernández, A., V erduzco- Mendoza, A., Hernández, E., Martínez-Burnes, J., & Whittaker , A. L. (2020). The Utility of Gri- mace Scales for Practical Pain Assessment in Labo- ratory Animals. In Animals (V ol. 10, Issue 10, 1838, p. 1-17). MDPI AG. https://doi.org/10.3390/ ani10101838 [48] National Research Council, Critical Needs for Research in V eterinary Science. (2005). Na- tional Academies Pr ess . pp. 1-239. https: //doi.org/10.17226/11366 [49] Navarro, E., Mainau, E., & Manteca, X. (2020). De- velopment of a Facial Expression Scale Using Far - rowing as a Model of Pain in Sows. In Animals (V ol. 10, Issue 1 1, 21 13, p. 1-1 1). MDPI AG. https:// doi.org/10.3390/ani10112113 [50] Nath, T ., Mathis, A., Chen, A. C., Patel, A., Bethge, M., & Mathis, M. W . (2019). Using DeepLabCut for 3D markerless pose estimation across species and behaviors. In Natur e Pr otocols (V ol. 14, No. 7, pp. 2152–2176). Springer Science and Business Media LLC. https://doi.org/10. 1038/s41596- 019- 0176- 0 [51] Neethirajan, S. (2021). Happy Cow or Thinking Pig? WUR W olf—Facial Coding Platform for Measuring Emotions in Farm Animals. In AI (V ol. 2, Issue 3, pp. 342–354). MDPI AG. https://doi.org/10. 3390/ai2030021 [52] Neethirajan, S., & Kemp, B. (2021). Digital Live- stock Farming. In Sensing and Bio-Sensing Resear ch (V ol. 32, No.100408, pp. 1-12). Elsevier BV . https: //doi.org/10.1016/j.sbsr.2021.100408 [53] Neethirajan, S., Reimert, I., & Kemp, B. (2021). Measuring Farm Animal Emotions—Sensor -Based Approaches. In Sensors (V ol. 21, Issue 2, 553, p. 1-22). MDPI AG. https://doi.org/10.3390/ s21020553 [54] Pandey , S., Kalwa, U., Kong, T ., Guo, B., Gauger , P . C., Peters, D. J., & Y oon, K.-J. (2021). Behav- ioral Monitoring T ool for Pig Farmers: Ear T ag Sen- sors, Machine Intelligence, and T echnology Adop- tion Roadmap. In Animals (V ol. 1 1, Issue 9, 2665, pp. 1-12). MDPI AG. https://doi.org/10.3390/ ani11092665 [55] Panksepp, J. (2005). Af fective consciousness: Core emotional feelings in animals and humans. In Con- sciousness and Cognition (V ol. 14, No. 1, pp. 30– 80). Elsevier BV . https://doi.org/10.1016/j. concog.2004.10.004 [56] Pennington, Z. T ., Dong, Z., Feng, Y ., V etere, L. M., Page-Harley , L., Shuman, T ., & Cai, D. J. (2019). ezT rack: An open-source video analysis pipeline for the investigation of animal behavior . In Scientific Re- ports (V ol. 9, No. 9:19979, p. 1-1 1). Springer Science and Business Media LLC. https://doi.org/10. 1038/s41598- 019- 56408- 9 [57] Pereira, T . D., Aldarondo, D. E., W illmore, L., Kislin, M., W ang, S. S.-H., Murthy , M., & Shae- vitz, J. W . (2018). Fast animal pose estimation using deep neural networks. In Natur e Methods (V ol. 16, Issue 1, pp. 1 17–125). Springer Science and Business Media LLC. https://doi.org/10. 1038/s41592- 018- 0234- 5 [58] Pessanha, F ., McLennan, K., & Mahmoud, M. (2020). T owards automatic monitoring of disease progression in sheep: A hierarchical model for sheep facial expressions analysis from video. In 2020 15th IEEE International Confer ence on Au- tomatic Face and Gestur e Recognition (FG 2020) . IEEE. p. 387-393. https://doi.org/10.1109/ fg47880.2020.00107 [59] Rashid, M., Silventoinen, A., Gleerup, K. B., & An- dersen, P . H. (2020). Equine Facial Action Coding System for determination of pain-related facial re- sponses in videos of horses. In U. G. Munderloh (Ed.), PLOS ONE (V ol. 15, Issue 1 1. e0231608, p. 1-18). Public Library of Science (PLoS). https://doi. org/10.1371/journal.pone.0231608 [60] Rashid, M., Broomé, S., Ask, K., Hernlund, E., Andersen, P .H., Kjellström, H., & Lee, Y .J. (2021). Equine Pain Behavior Classifi- cation via Self-Supervised Disentangled Pose Representation. in pr oc. of 2022 IEEE/CVF W inter Confer ence on Applications of Com- puter V ision (W ACV) , pp. 152-162. https: //doi.org/10.1109/WACV51458.2022.00023 [61] Raspa, F ., T arantola, M., Muca, E., Ber gero, D., Soglia, D., Cavallini, D., V ervuert, I., Bordin, C., De Palo, P ., & V alle, E. (2022). Does Feeding Manage- ment Make a Dif ference to Behavioural Activities and W elfare of Horses Reared for Meat Production? In Animals (V ol. 12, No. 14, 1740, p. 1-1 1). MDPI AG. https://doi.org/10.3390/ani12141740 48 Informatica 49 (2025) 33–48 S. Benedict et al. [62] Samadiani, Huang, Cai, Luo, Chi, Xiang, & He. (2019). A Review on Automatic Facial Expression Recognition Systems Assisted by Multimodal Sensor Data. In Sensors (V ol. 19, No. 8, 1863, p. 1-27). MDPI AG. https://doi.org/10.3390/s19081863 [63] Sénèque, E., Lesimple, C., Morisset, S., & Haus- ber ger , M. (2019). Could posture reflect welfare state? A study using geometric morphometrics in rid- ing school horses. In J. J. Loor (Ed.), PLOS ONE (V ol. 14, Issue 2, e021 1852, p. 1-20). Public Library of Science (PLoS). https://doi.org/10.1371/ journal.pone.0211852 [64] Singh, S., & Benedict, S. (2020). Indian Semi- Acted Facial Expression (iSAFE) Dataset for Hu- man Emotions Recognition. In Communications in Computer and Information Science (pp. 150–162). Springer Singapore. https://doi.org/10.1007/ 978- 981- 15- 4828- 4_13 [65] Soltis, J., Leong, K., & Savage, A. (2005). African elephant vocal communication II: rumble variation re- flects the individual identity and emotional state of callers. In Animal Behaviour (V ol. 70, No. 3, pp. 589– 599). Elsevier BV . https://doi.org/10.1016/j. anbehav.2004.11.016 [66] Soltis, J., Blowers, T . E., & Savage, A. (201 1). Measuring positive and negative af fect in the voiced sounds of African elephants (Loxodonta africana). In Journal of the Acoustical Society of America (V ol. 129, No. 2, pp. 1059-1066). Acoustical Society of America (ASA). https://doi.org/10.1121/ 1.3531798 [67] Sotocinal, S. G., Sor ge, R. E., Zaloum, A., T uttle, A. H., Martin, L. J., W ieskopf, J. S., Mapplebeck, J. C., W ei, P ., Zhan, S., Zhang, S., McDougall, J. J., King, O. D., & Mogil, J. S. (201 1). The Rat Gri- mace Scale: A Partially Automated Method for Quan- tifying Pain in the Laboratory Rat via Facial Expres- sions. In Molecular Pain (V ol. 7). SAGE Publications. https://doi.org/10.1186/1744- 8069- 7- 55 [68] Statham, P ., Hannuna, S., Jones, S., Campbell, N., Robert Colborne, G., Browne, W . J., Paul, E. S., & Mendl, M. (2020). Quantifying defence cascade responses as indicators of pig af fect and welfare using computer vision methods. In Scientific Re- ports (V ol. 10, No.8933, pp. 1-13). Springer Science and Business Media LLC. https://doi.org/10. 1038/s41598- 020- 65954- 6 [69] Stolz, K., Heyder , T ., Gloor , P . A., & Posegga, O. (2019). Measuring Human-Animal Interaction with Smartwatches: An Initial Experiment. In Stud- ies on Entr epr eneurship, Structural Change and In- dustrial Dynamics (pp. 165-182). Springer Inter - national Publishing. https://doi.org/10.1007/ 978- 3- 030- 17238- 1_10 [70] Thangavel, S., & Shokkalingam, C. S. (2021). The IoT based embedded system for the detection and dis- crimination of animals to avoid human–wildlife con- flict. In Journal of Ambient Intelligence and Human- ized Computing (V ol. 13, Issue 6, pp. 3065-3081). Springer Science and Business Media LLC. https: //doi.org/10.1007/s12652- 021- 03141- 9 [71] T omkins, S. S., & McCarter , R. (1964). What and Where are the Primary Af fects? Some Evidence for a Theory . In Per ceptual and Motor Skills (V ol. 18, Issue 1, pp. 1 19-158). SAGE Publications. https: //doi.org/10.2466/pms.1964.18.1.119 [72] T sai, M.-F ., & Huang, J.-Y . (2021). Sentiment analy- sis of pets using deep learning technologies in artifi- cial intelligence of things system. In Soft Computing (V ol. 25, Issue 21, pp. 13741-13752). Springer Sci- ence and Business Media LLC. https://doi.org/ 10.1007/s00500- 021- 06038- z [73] W aller , B. M., Julle-Daniere, E., & Micheletta, J. (2020). Measuring the evolution of facial ‘expres- sion’ using multi-species F ACS. In Neur oscience; Biobehavioral Reviews (V ol. 1 13, pp. 1-1 1). Elsevier BV . https://doi.org/10.1016/j.neubiorev. 2020.02.031 [74] Xu, B., W ang, W ., Guo, L., Chen, G., Li, Y ., Cao, Z., & W u, S. (2022). CattleFaceNet: A cattle face identi- fication approach based on RetinaFace and ArcFace loss. In Computers and Electr onics in Agricultur e (V ol. 193, No. 106675, p. 1-10). Elsevier BV . https: //doi.org/10.1016/j.compag.2021.106675 [75] Y u-Ju Chiu, Shajulin Benedict, and Michael Gerndt. (2023). E2D2: Elephant Emotion and Distraction Detection Framework using Edge-Enabled YOLOv5 Deep Learning Algorithm. in 2023 IEEE Inter - national Confer ence on Contemporary Comput- ing and Communications (InC4) , Bangalore, In- dia, p. 1–6, https:doi.org/10.1109/InC457730. 2023.10262860 [76] Zeppelzauer , M., & Stoeger , A. S. (2015). Es- tablishing the fundamentals for an elephant early warning and monitoring system. In BMC Resear ch Notes (V ol. 8, Issue 409, p. 1-15). Springer Science and Business Media LLC. https://doi.org/10. 1186/s13104- 015- 1370- y [77] Zhu, H., Salgırlı, Y ., Can, P ., Atılgan, D., & Salah, A. A. (2023). V ideo-based estimation of pain indicators in dogs. pr oc. of 1 1th International Confer ence on Af- fective Computing and Intelligent Interaction (ACII) , pp. 1-8. https://doi.org/10.1109/ACII59096. 2023.10388142