RTC-DETR234: Real-Time CBAM Detection Transformer 234—Diagnostic Detection of Band Sawtooth and Fractures

In industrial production processes, accurate and fast detection of band sawtooth deformations is of critical importance. Traditional detection approaches are not sustainable at industrial scale due to their high human attention requirements and low efficiency. In this study, we present RTC-DETR234, a task-specific configuration of the RT-DETR architecture designed for detecting small and low-contrast band sawtooth fractures. The proposed model uses the P2–P4 scales to provide higher-resolution feature extraction instead of the P3–P5 scaling structure of the baseline RT-DETR model. Furthermore, CBAM modules are integrated into the HGBlock-based backbone to refine channel- and spatial-level feature responses associated with small and low-contrast fracture regions. The proposed model, the baseline YOLO models, and the baseline RT-DETR model are trained and compared on a dataset consisting of band sawtooth defects. The proposed model achieved an average mAP of 0.9938 for tooth detection and 0.9178 for fracture detection across multiple independent runs.