TCCGAN: A Stacked Generative Adversarial Network for Clinical Tongue Images Color Correction

Tongue diagnosis has become an important part of Stomatology. However, the light source environment, camera equipment, and camera settings will impact the acquired tongue image's quality in the actual scene. To solve the problem of color constancy in tongue image acquisition, we propose the TCCGAN network to correct the tongue image's color. We first present a differentiable weighted histogram network for color feature extraction, which is used in a new upsample module called the mixed feature attention upsample module to assist image generation. Then, a stacked network is built to generate tongue images from coarse to fine. Finally, we analyze the limitations of the traditional loss function and propose a new loss function. The experimental results show that the image quality generated by our method is better than other methods, and the accuracy of the downstream diagnosis and classification task is significantly improved.