Attention Mechanism-Based Deep Learning Denoising of Scanned Point Cloud for Rocket Tank Panel

Point cloud denoising plays a vital role in the geometric quality measurement of rocket tank panels. However, it is difficult to remove the large-scale noise points accurately from the collected scanned point cloud of the rocket tank panel, which poses a challenge to point cloud denoising. Therefore, an attention mechanism-based deep learning network, called PointAPL, is proposed for point cloud denoising. To deal with the characteristics of large-scale noise, the attention pooling layer (APL) is designed to append on the top of dilated PCPNet to enhance the global feature extraction performance of the point cloud through this attention mechanism. Meanwhile, in order to improve the network training efficiency, the adaptive weight cross-entropy (AWCE) loss function is proposed. The trained network can automatically identify the object points and noise points from the raw scanned point cloud. A set of clean points is then generated for high-precision geometric quality measurement of the rocket tank panel. Besides, extensive experiments demonstrate that the proposed method outperforms the traditional denoising methods in terms of both qualitative and quantitative evaluation metrics.