A 3D Reconstruction of Binocular Endoscope Based on Self-supervised Dilated Convolution

With the development of endoscopy-related technologies, minimally invasive surgery is still difficult to operate and requires doctors to have many years of operating experience. A disparity estimation method based on neural network is proposed. The end-to-end disparity estimation method can combine image information from different range sizes to improve the robustness of the network to the interference of specular reflection and single texture of illumination. In order to utilize the information in a wider range around the pixel, this model uses multiple dilated convolutions with different dilation rates to extract the features of the same hidden layer and then fuse them. This enhances the network receptive field and improves the network's ability to perceive the depth of detail. At the same time, in order to compare the similarity of the left and right images, this paper proposes a parallax regression model using the pseudo-twin structure, which is easier to converge than the model training of the twin structure. Experiments show that this method does not greatly improve the network parameters while improving the ability of the network to resist light interference.