A binocular stereo matching method based on a convolutional neural network

Abstract

The invention provides a binocular stereo matching method based on a convolutional neural network. By improving the existing DispNet network model used for disparity estimation, the subpixel convolution is used to replace the up-sampling layer in the original network model. Compared with the original upsampling operation on a higher resolution image, sub-pixel convolution is directly used in low-resolution images, which not only improves the computational efficiency, but also accelerates the matching speed for the whole network. At the same time, the method improves the good performance of thenetwork model, adds a wealth of detail information, and solves the problem that ill-conditioned regions can not be correctly matched.

Description

Technical field [0001] The invention belongs to the technical field of computer vision, and specifically relates to a binocular stereo matching method based on a convolutional neural network. Background technique [0002] Stereo matching is usually described as an optimization problem that can be divided into several stages. Until recent years, with the development of convolutional neural networks, it can be described as a learning task. Using a large amount of existing data for training, the speed and accuracy of matching convolutional neural networks are better than traditional methods. At present, there are many stereo matching methods through convolutional neural networks, which are mainly divided into three categories: matching cost learning, regularization learning and end-to-end parallax learning. Matching cost learning uses different training sample sets for training, but does not pay attention to the imbalance between sample sets. Although data-driven similarity measu...

AI Technical Summary

Problems solved by technology

These problems will cause the matching failure

Images