Blind image quality metric for color images based on human vision system and deep CNN

In this study, a blind image quality metric (BIQM) for human vision system (HVS) has been developed by using deep convolutional neural network (CNN) for color images. The proposed quality metric consists of four stages. In the first stage, the colored image is transformed into YUV color space due to its compatibility with the HVS. In the second stage, pre-trained deep CNN is defined according to the noise type in the image (jpeg, gaussian, etc.). In the next step, the image whose quality is to be measured is denoised with deep CNN, and the image obtained as a result of this process is assumed to be the reference image. Thus, the problem is transformed into a full-reference quality measurement problem. In the last stage, each color channel (i.e., Y, U, and V) of the image whose quality is to be measured with the obtained reference image is measured by structural similarity (SSIM) method. The measurement results are weighted by considering the HVS. The experimental results show that the developed metric was able to sort the images that other quality metrics could not sort. Moreover, a survey was conducted with 100 participants to rank images in 3 different qualities, and the survey results support that the proposed method is compatible with the HVS. It also modified the SSIM which is normally a full-referenced method, to use as no-reference. Moreover, thanks to the proposed BIQM, more appropriate results are obtained compared to commonly used no-reference quality metrics. MATLAB codes and some images used for the proposed BIQM are available in the link: https://bit.ly/3NIrNBs