STD2: Swin Transformer-Based Defect Detector for Surface Anomaly Detection

• Published in: IEEE transactions on instrumentation and measurement p. 1
• Main Authors: Mia, Md Sohag, Li, Chunbiao
• Format: Journal Article
• Language: English
• Published: IEEE 06.11.2024
• Subjects: Adaptation models; CIoU; Computational modeling; Computer architecture; Computer vision; Deep learning; Defect detection; Deformable Pooling; Faster R-CNN; Feature extraction; Inspection; PAFPN; Steel; Swin Transformer; Transformers
• ISSN: 0018-9456; 1557-9662
• DOI: 10.1109/TIM.2024.3492728

In the quest for manufacturing excellence, this paper presents a pioneering approach to defect detection on steel surfaces. Complex and small defects on steel surfaces manifest in diverse forms, such as irregular shapes and varying sizes, making it challenging to devise a singular detection model capable of effectively detecting these multifarious defect types. Yet, another challenge lies in the close similarity between defects and non-defective components. Additionally, the fluctuating environmental factors exposed to steel structures introduce noise and artifacts into image data, compounding the complexity of precise defect detection. To address these issues, we initially employed the Swin Transformer as the core element of Faster R-CNN. Secondly, we ingeniously integrated a Path Aggregation Feature Pyramid Network (PAFPN) into the architecture. This allows us to capture a comprehensive array of defect feature maps and enhance the network's capability to identify defects at various scales. After that, we swapped out the RoI Pooling for the more sophisticated Deformable Pooling method to get more precise defect localization information. Finally, the standard IoU loss is substituted with Complete Intersection over Union (CIoU) to address issues of duplicate proposals and improve overall performance by taking into account the geometry of the bounding boxes in addition to their overlaps. Our proposed STD2 has been rigorously evaluated on diverse steel surface defect datasets, including NEU-DET (yielding an impressive mAP of 81.05%), GC10-DET (with a notable mAP score of 72.38%), and PCB (achieving an outstanding mAP of 99.00%), which demonstrate its exceptional performance in surface anomaly detection. Code is available at https://github.com/Shuvo001/STD2.