The Research Focuses on Enhancing the RTDETR Model for Improved Detection of PCB Defects

In response to the low accuracy, slow speed, and compatibility issues commonly exist in current PCB circuit board defect detection methods, this paper proposes an improved circuit board defect detection model called RTDETR-CGRFPN based on RTDETR. This model is designed with a Rectangular Self-Calibration Module for spatial feature reconstruction and pyramid context extraction. Additionally, considering the limited background information of PCB circuit boards and the performance limitations of general object detection models, FuseBlock Multi and Dynamic Interpolation Fusion modules are added to the original model for multi-scale feature fusion through dynamic interpolation. Furthermore, dynamic interpolation and multi-feature fusion are utilized to further enhance target detection capabilities in complex backgrounds. Moreover, Pyramid Context Extraction module is also introduced to enhance the model's perception of context and improve its ability to detect small defects. Experimental results demonstrate that our proposed algorithm achieves an average accuracy rate of 98.1% with a recall rate of 99.1% on the test dataset. In comparison, the original RTDETR-r18 algorithm has an average accuracy rate of 95.0% with a recall rate of 90.9%. Clearly, our proposed algorithm outperforms in these two metrics significantly. Additionally, this model maintains relatively few parameters while ensuring performance improvement and computational efficiency simultaneously; thus enabling quick and accurate defect detection in PCB circuit boards potentially providing technical support for fields such as smart factories.