Hopfield Neural Network based Stereo Matching Algorithm

Traditional feature-based stereo matching methods fail to obtain dense disparity map, and in this paper we directly treat each pixel point as a matching primitive, treating the stereo-matching problem as a combinatorial optimization problem of finding labels for each pixel point First, an energy function is introduced to represent the matching constraint cost function for the stereo matching. Then, based on the energy minimization algorithm, a three-dimensional matching algorithm based on a two-dimensional Hopfield neural network is proposed that considers the polar constraint, uniqueness constraint, similarity constraint, and smoothness constraint, mapping the variability directly to the neural state space and introducing these constraints into the energy function through the neural state. Finally, iterations are used to ensure that the energy function converges to a local minimum. In addition, constructing a Hopfield neural network on each epipolar line, based on the epipolar constraint, not only reduces the number of neurons but also increases the computational speed. The experimental construction was performed on Middlebury and the results showed good convergence of the energy function and good accuracy of the matching results.