Dual Attention CycleGAN for Single-Depth Estimation in Digestive Endoscopy

Abstract: In response to the problems of edge information blur and poor local interference information processing capability in the digestive endoscopy diagnosis and treatment system, this paper proposes a single-depth estimation method based on dual attention cycle generative adversarial networks (DA-CycleGAN) to achieve accurate estimation of digestive depth information. By utilizing the global attention module to improve network accuracy and reduce information loss through global cross-dimensional interaction, as well as using the channel attention module to enhance the correlation between channels and reduce local interference information in endoscopic images, the discriminator adopts multi-scale feature fusion to improve discrimination ability and balance the performance of the generator and discriminator. The results show that the proposed method can predict good results in digestive endoscopy scenes, with an average accuracy improvement of 1.74%, 10.84%, and 0.69% for gastric, small intestinal, and colon datasets, respectively, compared to other unsupervised methods.