Creating a Neural Network Noise Reduction Metric Based on Human Color Perception

The presence of noise in images is one of the most significant challenges in modern computational photography, which explains the rapid and continuous development of this field. One of the most complex issues is finding the optimal image quality assessment (IQA) method. Currently, both analytical IQA metrics, such as PSNR and SSIM, and neural network-based metrics, such as LPIPS and CLIP-IQA, are in use. Interestingly, no existing IQA methods take into account knowledge of the color perception characteristics of the standard CIE XYZ observer. This work is the first to demonstrate that incorporating such knowledge using the CIEDE2000 color difference formula and Euclidean distance in the CIELUV space can improve the accuracy of image quality assessment. We propose a new IQA metric based on a combination of classical and color-based formulas for comparing images. Experiments on a new dataset show an increase in correlation with human opinion scores by up to 20%, confirming the high potential for further development in this direction.