Rapid Batch Three-Dimensional Reconstruction of Point Clouds Based on Multi-Label Classification

Existing traditional three-dimensional (3D) reconstruction methods focus on preserving surface details of 3D objects; however, they cannot realize rapid reconstruction of 3D objects but can only reflect the category, shape, and other typical features of 3D objects. To tackle this problem, a rapid batch method for 3D object reconstruction is proposed. First, point clouds are simplified before they are denoised; an enhanced k-neighbor denoising algorithm with double-threshold constraints is proposed. The denoising performance of this algorithm is compared with that of two traditional denoising methods. Second, the categories of 3D point clouds are obtained using the dynamic graph convolutional neural network(DGCNN). Finally, after matching each category with the categories in a pre-built 3D model library, corresponding 3D models are called sequentially to achieve the batch rapid 3D reconstruction. If there is no corresponding category in the pre-built 3D model library, the 3D point clouds of the corresponding category can be acquired and added to the DGCNN for training and evaluating. The 3D reconstruction result of each category is added to the pre-built 3D model library to increase the number of categories in the library. The effectiveness of our proposed method is verified using the incremental ModelNet40 model. The experimental result shows that the 3D reconstruction method incurs the cost of 11.98 s for 120 point clouds of 100000 points, 29.06 s for 120 point clouds of 700000 points, and 109.98 s for 120 point clouds of 1200000 points, which is approximately 10 times faster than that of the traditional method considered in this study. Overall, this method can significantly improve the efficiency of 3D reconstruction for point clouds with both large order of magnitude and a large number, as well as realize real-time batch rapid 3D reconstruction of point clouds.