Superpixel-based regional anomaly detection for LED chips using local statistical features

The quality control of light-emitting diode (LED) chip constitutes a pivotal stage in the production process. The utilization of machine vision provides an effective methodology for the identification of defects in LED chips. However, the diverse visual characteristics of chip defects, in conjunction with the limited amount of available defect data present a significant challenge to the effective detection of chip defects. To address this problem, a superpixel-based regional anomaly detection method is proposed using local statistical features (SRAD-LSF) for diverse LED chips. Initially, an improved superpixel segmentation method is proposed based on the reconstructed colour channel, with the objective of recalling accurate boundary details and providing higher-level statistical information. Subsequently, a chip image feature extraction is proposed using local statistical information, including colour and position features in an efficient manner. Based on the extracted local statistical features, an unsupervised region segmentation and anomaly detection methods are proposed for diverse LED chips. The proposed regional anomaly detection is capable of achieving accurate detection performance without the necessity of negative training samples. The experimental results on the constructed chip image dataset demonstrate that the proposed method can accurately detect typical chip anomalies. The precision and recall of SRAD-LSF achieve 96.36 % and 92.80 %, respectively.