Multi-view Stereoscopic 3D Reconstruction Based on Adaptive Feature Fusion

Aiming at the problems of substation equipment such as excessive outdoor noise and uncontrollable illumination that lead to poor 3D reconstruction model, this paper studies the CasMVSNet 3D reconstruction algorithm, and proposes an adaptive feature body fusion module, which can adapt to different lighting information and improve model reconstruction. In addition, a depth information correction module based on attention mechanism is proposed, which can make the reconstructed point cloud coordinates more accurate. Firstly, multi-level feature extraction is performed on the image based on the FPN structure. Then, the feature volume is obtained by the differentiable homography transformation; Then, the proposed adaptive feature volume fusion module constructs the cost volume, and the cost volume performs regularization operation. Finally, the image noise is eliminated by the correction module. Testing on the DTU public dataset, the method has improved both accuracy and completeness, with an accuracy of 0.3109 and an integrity of 0.4568, and a qualitative comparison using the multi-view image set of the substation equipment has a higher completeness of the method. Through experimental comparisons, the network structure proposed in this paper has improved the accuracy and integrity.