Stereo Matching Method Based on Lightweight PASMnet Neural Network
A lightweight PASMnet neural network for stereo matching is proposed for the problems of poor results of traditional methods for stereo matching and 3D terrain reconstruction, and large models and low efficiencies of deep learning methods. First, structured pruning is performed on PASMnet to refine the parameter configuration, reduce the number of model parameters, and improve the model running speed. Then the normalized image is polar line corrected to obtain a stereo image with parallel imaging planes for the stereo matching task. Validated on the KITTI2015 dataset, the experimental results show that the improved model parameter quantity is reduced to 1/6 of the original model, and the running time is 0.56 s. The pruned and quantized model runs faster, has shorter in-ference time, and can be used for stereo matching of lunar surface images, which extend the model application.
万方数据
维普
