Stereo Matching with Adaptive Disparity Search Based on Multi Scale Features

Stereo matching, a core research direction in binocular vision, has garnered significant attention for its cost-effectiveness in acquiring depth information. In neurosurgery, stereo matching technology is actively utilized to enhance surgical precision and improve treatment outcomes. However, acquiring high-precision depth information is hindered by the limited availability of datasets and the high computational complexity. To address these challenges, this study constructs a virtual dataset of neurosurgical scenes consisting of 2000 pairs of binocular images, along with complete depth maps, segmentation masks, and optical flow data, supporting multitask learning. Based on this, we propose a multi-scale, feature-driven adaptive disparity range search algorithm. The algorithm expands the receptive field of feature maps to capture mid-to-long-range features and enhances feature representation using a multi-branch structure. Additionally, depthwise separable convolution is introduced to reduce computational load and ensure real-time performance, while adaptive adjustment of the disparity search range significantly mitigates mismatches in pathological regions. Furthermore, cost aggregation and multi-branch parallel computation are employed to further enhance computational efficiency. Extensive experimental results demonstrate that the proposed method outperforms existing baseline models in terms of accuracy while maintaining superior real-time performance.