Improved Fast Stereo Matching Algorithm Based on Convolutional Neural Networks

Various stereo matching algorithms are widely used in many fields nowadays. In order to solve low speed problem in convolutional neural network for stereo matching algorithms (MC-CNN-fst), an improved stereo matching algorithm based on convolutional neural network was proposed. By modifying pooling layers and normalization layer, the structure of the convolutional neural network has been reformed. Among them, different pooling layers are introduced to retain more image information, and the normalization layer is introduced to accelerate the speed of convergence of training, and finally the disparity map is obtained through subsequent processing steps. To evaluate the performance of the proposed algorithm and the original one under the same experimental environment. We conduct the experiment on two public datasets: KITTI and Middlebury. Experimental results demonstrate that the refinement of the network structure boosts the speed for the feature extraction procedure by 50%, and the total operation time is reduced by 12.5%.