Binocular speckle structured light disparity estimation based on spatial pyramid densely connected stereo matching network

Stereo matching with speckle structured light is a crucial task in the field of optical 3D measurement and computer vision. Currently, stereo matching networks for structured light images still struggle with limited accuracy and lack of robustness particularly in term of details and edges. In order to improve the disparity estimation for speckle structured light images, in this paper we propose a spatial pyramid densely connected stereo matching network(DSSMNet). The proposed network incorporates a hybrid module that simultaneously performs spatial pyramid pooling and densely connected operations, and integrates the Convolutional Block Attention Module (CBAM) to enable multi-scale feature extraction. In addition, a new loss function is designed to optimize the edge information in network training. The results from simulated and real datasets demonstrate that, the network proposed in this paper can predict disparity results with more refined and intricate details and edges, and outperforms the state-of-the-art networks such as PSMNet, GWCNet, ACVNet and Lac-GwcNet in visual quality and in quantitative metrics.