Solving the Stereo Matching in Disparity Image Space Using the Hopfield Neural Network

As the accuracy of boundary region is the bottleneck of stereo matching, here a feature-based matching algorithm is used to focus on the matching problem in the boundary. Firstly, for the stereo matching, we propose a new boundary feature extraction algorithm based on the RBF, and make the boundary as the region to be matched. After considering various constraints the matching needs to meet, we build a new energy function, and then use the Hopfield neural network (HNN) to do the optimization. Because the object area is all of the boundary, so it is not wise to input too many neurons to the network. To further reduce the complexity of the algorithm, it is proposed to construct a network model from the disparity image space, which is an innovative approach. Finally, a large number of experiments are conducted to validate the algorithm, including the standard images, the noise images and the real scene images. Experimental results show that the new algorithm can greatly improve the accuracy of the boundary so as to overcome the bottleneck. At the same time, the accuracy of the whole region will be also significantly improved. When tested in complex cases, the algorithm can still get a good result, so it can be said that the robustness and performance of the algorithm are quite good.