A Two-Stage Detection Method for Tiny Defects in Transparent Components Based on Pseudolabel Generation

Defect detection in transparent components is important for improving their quality. However, such defects are typically tiny and diverse, occurring at different scales, which makes manual labeling difficult and time-consuming. Thus, this study proposes a data acquisition system based on light sensors and a two-stage detection method for tiny defects in transparent components using pseudo-label generation. The first stage consists of an unsupervised detection model called PatchCore-PBA and a pseudo-label mining strategy. PatchCore-PBA contains a novel feature extraction module called PBA-Net, which is composed of a composite convolutional layer and a multi-scale integration layer. This module enhances PatchCore’s feature extraction capability. Subsequently, a pseudo-label mining strategy based on ATS-RPN is formulated to obtain high-quality pseudo-labels. The second stage proposes a new detection model, YOLO-ML. YOLO-ML uses an ML-Net, which is composed of a multi-scale wavelet convolution module and a variable large kernel attention module. These modules enhance the expression of features of tiny-scale defects. This improves the performance of YOLO-ML in detecting multi-scale defects. Furthermore, the study establishes a new transparent component dataset, LENS-DET, to provide a benchmark for transparent component defect detection. Compared to other defect detection methods, the proposed method requires no manual labeling and has advantages in terms of both detection accuracy and speed. The dataset is available at https://github.com/RoomforSds/LENS-DET.