A General Endoscopic Image Enhancement Method Based on Pre-trained Generative Adversarial Networks

Endoscopic images frequently have image quality problems due to the limitations of surgical instruments and the impact of surgical operations, such as uneven illumination, smogginess and color deviation. For deep learning based on enhancement methods, independent training lacks sufficient defect images and generalization capability, and combined training with mixture of data cannot identify diverse specific tasks. To address these issues, we propose a general method based on pre-trained generative adversarial network with a specified transfer learning strategy to obtain high-quality images. Initially, we independently train a standard network based on a universal task, e.g., uneven illumination, where a pre-trained model is extracted as a backbone with partially shared generator. Then, we transfer the backbone to more potential image enhancement tasks. Experiments on uneven illumination, smogginess, and color deviation indicate that the model successfully shares common features of high-quality images and responds specifically to different defects as well.