A multi-view 3D reconstruction method that integrates patchmatch and efficient channel attention cascades

Multi-view 3D reconstruction refers to reconstructing the 3D structure and dynamic changes of the real world from images captured at different viewpoints. It is critical for in-depth analysis and understanding of real-world scenes. To address issues like insufficient feature extraction capabilities and weak correlation between feature extraction and depth estimation in existing multi-view 3D reconstruction methods, this paper proposes a multi-view 3D reconstruction approach integrating block matching and cascaded effective channel attention mechanisms. First, standard convolutions in the feature pyramids are replaced with grouped ghost convolutions in the feature extraction module, enabling the model to better capture subtle feature changes and improve its capability on complex tasks. Second, block matching networks at different scales are cascaded with the effective channel attention mechanism in the depth estimation layer to further enhance the network's learning ability. Finally, experimental results demonstrate that compared to the baseline network, the proposed algorithm improves point cloud reconstruction accuracy and overall performance on the DTU dataset by 5.9% and 1.9% respectively. It also shows competitive advantages on the custom dataset. This validates that the proposed network achieves high reconstruction accuracy and can be applied to practical high-precision 3D reconstruction tasks.