Low-Light Gastroscopic Image Enhancement Network Based on CycleGAN

Aiming at problems such as uneven illumination and low image contrast in gastroscopic images, an unsupervised low-light gastroscopic image enhancement network based on CycleGAN is proposed. Firstly, a down-sampling module based on wavelet transform is designed to reduce the loss of detail information during the generator's down-sampling process. Secondly, a dynamic feature fusion module is used as a skip connection structure to enhance the fusion of multi-scale features. Finally, an illumination-aware attention module is introduced to improve the accuracy of illumination estimation and suppress local overexposure. This method was compared with RRDNet, EnlightenGAN, LE_GAN, and SCI. It showed excellent performance in both the Natural Image Quality Evaluator (NIQE) and Blind/Referenceless Image Spatial Quality Evaluator (BRISQUE) metrics. Overall, this method can effectively enhance the brightness of gastroscopic images while preserving texture information, optimizing the visual recognition of mucosal vessel textures and early lesions, which helps improve the accuracy of clinical diagnoses and holds significant clinical application value.