Geometry-Enhanced Multi-Level DynamicDisparity Estimation and Virtual View Synthesis Method for Stereo Vision

Accurate disparity map estimation is crucial for computer vision tasks such as virtual view generation and 3D modeling. Addressing the issues of low disparity estimation accuracy in textureless, repetitive-texture, and occluded regions, as well as artifacts in generated views observed in existing methods, this paper proposes a geometry-featureenhanced multi-level dynamic disparity estimation and virtual view generation method. The approach extracts multi-scale regional features from the source image using dilated convolutions with differentiated dilation rates and incorporates a Feature Pyramid Network (FPN) along with a multi-scale geometric enhancement module to capture contextual and edge information from the reference image. During the fusion stage, a gradient saliency attention mechanism is introduced to enhance responses in critical regions, and a 3D geometric encoding volume is constructed. Subsequently, a three-stage lightweight Gated Recurrent Unit (GRU) network is employed to iteratively refine the initial disparity map, with knowledge distillation incorporated to accelerate training convergence. Finally, a context-guided spatial upsampling strategy is utilized to output a highresolution disparity map, and high-quality virtual views are generated through geometric point cloud mapping and viewpoint reconstruction. Experimental results demonstrate that the proposed method outperforms existing mainstream approaches in terms of disparity estimation accuracy, robustness, and virtual view quality, exhibiting significant potential for broad engineering applications.