Abstract
Anomaly detection is pivotal in various industrial areas, including automotive, aircraft, and manufacturing. The primary goal of anomaly detection is to minimize human intervention while enhancing detection accuracy. Recently, researchers have achieved considerable progress in pinpointing defects using image-based techniques. However, these methods present two main challenges: (i) the detection accuracy is often influenced by various image factors such as perspective, lighting, and color variations, and (ii) some defects are subtle and challenging to discern in images. Recognizing these constraints shows a growing interest in exploiting three-dimensional (3D) point cloud data. These data offer a detailed and accurate representation of an object’s surface, promising better localization of surface defects. However, the 3D point cloud processing has its challenges. Two key issues include: (i) point cloud-based methods need to be both transformation and permutation invariant, as different transformations or permutations of the 3D point cloud can describe the same object structures, and (ii) at times, anomalies might share similar geometric features (e.g., local surface variations, etc.) with a non-defective surface, especially for objects with a complex shape, which makes it hard to distinguish the normal and abnormal surfaces in the space of geometric features. This paper introduces a novel, lightweight, unsupervised approach to detect and localize surface defects from 3D point cloud data, emulating human recognition capabilities and excelling in localization accuracy. We compare the performance of the proposed method and two classical unsupervised methods, K-means clustering based on cosine similarity between normal vectors and threshold method based on local surface variations on the real scanned dataset. The results show that our proposed method receives the best defect localization accuracy compared with baseline methods.