In the production process of vehicle stamped parts, local deviations from the nominal geometry of stamped metal sheet, i.e. defects, may occur. The recognizability of these defects is affected by illuminations, and image-based anomaly detection methods cannot effectively detect abnormal vehicle stamped parts. To solve the effect of illuminations, this paper takes the three-dimensional (3D) point cloud of stamped parts as the research object, and proposes a Multiscale Point feature-based 3D anomaly detection method (MP3D). To extract multiscale point features, this paper proposes a local aggregation module. Local aggregation module realizes feature aggregation of disordered points, and the aggregated point features have a larger receptive field. Features of different receptive fields are aggregated for multiscale anomaly detection. In addition, this paper designs a 3D anomaly generation strategy, which generates diverse abnormal samples by constructing local defects. Since the anomaly detection task requires classifying every point of the sample, there is an imbalance in the number of normal points and abnormal points. This paper improves the cross entropy loss for the anomaly detection task. To evaluate the performance of the proposed MP3D, this paper conducts extensive experiments on the MVTec 3D Anomaly Detection (MVTec3D-AD) dataset and a real stamped part dataset. Experimental results demonstrate that MP3D achieves effective anomaly detection performance at both the sample and point levels.
扫码关注我们
求助内容:
应助结果提醒方式:
