3D point cloud reconstruction and defect detection of mechanical parts based on recognition feedback-optimized VoteNet

Abstract

The purpose of this study is to provide more reliable technical support for industrial quality inspection. To tackle the limitations of inadequate detection precision and low levels of automation inherent in conventional quality inspection processes for mechanical components, this study presents a framework for three-dimensional (3D) point cloud reconstruction and defect detection that leverages recognition-driven feedback mechanisms. Based on binocular structured light technology and an improved VoteNet model, a complete process from 3D data acquisition to defect identification is built. First, the imaging principle and calibration methodology of the pinhole camera model are investigated, and a high-contrast calibration target featuring white circular markers on a black background is adopted to perform binocular system calibration. A least-squares ellipse fitting algorithm is employed to pinpoint the centroid of each marker, which enhances calibration precision and exhibits greater robustness against minor fluctuations in ambient lighting typical of industrial environments. Then, the 3D point cloud reconstruction is carried out. The structured light projector and the camera cooperate to collect multi-view images. After statistical filtering and radius filtering, noise is removed, and then the occluded area is interpolated to complete, and finally, high-quality point cloud data is output. In defect detection, the VoteNet model is improved for the disorder and sparseness of point clouds, and the symmetric function is introduced to optimize feature coding, and a three-layer multi-scale feature fusion structure is designed. The experiment is based on the MachPart-Defects data set, and the results look ideal: the maximum average measurement error after multi-view point cloud precise registration is only 0.0556 mm, and the maximum average relative measurement error is as low as 0.95%. The overall F1 value of the improved VoteNet model reaches 92.4%, which is better than the mainstream algorithms such as dilate gated convolutional neural network, and the F1 value of detecting tiny defects below 0.1 mm is 86.9%. Especially in the detection of defects such as cracks and burrs. Even in the complex scene where the surface of the part is highly reflective, the detection accuracy of the model can be maintained above 82%. The technical process constructed in this study can basically meet the requirements of industrial quality inspection in terms of accuracy and stability, and can provide a reference for quality control of the machinery manufacturing industry.