Ogrodje za ocenjevanje barve svetila z linearnimi digitalnimi filtri, prilagojenimi na posamezno sliko : doktorska disertacija

slovenian

doktorska dela

2020

colour constancy   convolution   filter adaptation   grey edge framework   ICA   illuminant estimation   independent component analysis   PCA   principal component analysis

Univerza v Mariboru, Fakulteta za elektrotehniko računalništvo in informatiko

A problem of colour constancy or a problem of estimating the colour of scene%s uniform illuminant based on captured digital images, respectively, is researched in this doctoral thesis. A generalisation of the established greyedge framework is presented by introducing a discretization and convolution operation. Filters adapted to image content are proposed for estimating the illuminant with our framework. Filters are constructed using either Principal Component Analysis (PCA) or Independent Component Analysis (ICA). The method is sped up so that instead of calculating the whole convolution only samples from the image are used or subsampling technique is applied, respectively. The constructed adapted filters are also used as individual colour constancy algorithms in a combinational method. Three major experiments on established public evaluation data sets Gegler-Shi and NUS are conducted in the experimental part of this dissertation. Results of the proposed methods are in the first experiment compared to results of the grey-edge framework methods and of other modern illuminant estimation methods. The proposed colour constancy methods from our framework achieve lower errors and lower mean rank than other grey-edge framework methods. The proposed combinational method achieves lower errors than the original method. The robustness of our colour constancy methods and grey-edge framework methods are in the second experiment compared with respect to minor scene changes. The changes are simulated by image rotation and translation. It is demonstrated that the proposed colour constancy method is robust to scene rotation when filters adapted to image content by Principal Component Analysis are used in our framework. On the other hand all compared methods are sensitive to scene translations. Accuracy of the illumination estimation is compared in the third experiment with respect to our proposed framework if either subsampling technique is used or the illuminant is estimated by convolution. A statistically significant difference in results is shown between both approaches. The illuminant estimation obtained by subsampling is more accurate than the illuminant estimation calculated by convolution comparable or better results at lower computational complexity than other modern colour constancy methods.

07.08.2023