Abstract
In minimally invasive surgery, endoscopic image quality plays a crucial role in surgery. Aiming at the lack of a real parallax in binocular endoscopic images, this article proposes an unsupervised adaptive neural network. The network combines adaptive smoke removal, depth estimation of binocular endoscopic images, and the 3D display of high-quality endoscopic images. We simulated the smoke generated during surgery by artificially adding fog. The training images of U-Net fused by Laplacian pyramid are introduced to improve the network's ability to extract intermediate features. We introduce Convolutional Block Attention Module to obtain the optimal parameters of each layer of the network. We utilized the disparity transformation relationship between left- and right-eye images to combine the left-eye images with disparity in HS-Resnet to obtain virtual right-eye images as labels for self-supervised training. This method extracts and fuses the parallax images at different scale levels of the decoder, making the generated parallax images more complete and smoother. A large number of experimental research results show that the scheme can remove the smoke generated during the operation, effectively reconstruct the 3D image of the tissue structure of the binocular endoscope, and at the same time, preserve the contour, edge, detail, and texture of the blood vessels in the medical image. Compared with the existing similar schemes, various indicators have been greatly improved. It has good clinical application prospects.