Stereo matching method based on convolutional neural network

According to the invention, a deep learning method is mainly adopted to process an input stereo picture pair so as to obtain a disparity map, so that the problem of precision reduction of a stereo matching network in a repeated texture region is at least partially solved. Firstly, a twin network is constructed by using residual blocks, feature extraction is performed on a stereo picture pair, then similarity of pixel points is calculated by using geometric priori, and a three-dimensional cost volume of geometric guidance is constructed. And predicting an initial disparity map by using an improved Soft Argmax function, and predicting a disparity residual error by using a correlation pyramid and an update operator for multiple times so as to obtain accurate disparity estimation. According to the method, the problem that accurate matching points cannot be found in ill-conditioned areas such as repeated textures is partially solved, and a good balance can be made between the computing power and the precision. 本发明