Surface defect detection algorithm for special-shaped electronic components based on improved RT-DETR

Abstract

The detection of surface defects of special-shaped electronic components is a key link to improve the level of insertion technology of special-shaped plug-in machines. Traditional manual detection is susceptible to human subjectivity, the efficiency of template matching algorithm is low, and in the case of insufficient sample data, the existing deep learning technology has problems such as low accuracy and insufficient real-time performance in defect detection. In order to improve the accuracy and real-time performance of special-shaped component detection, this paper improves the object detection model RT-DETR and proposes a real-time multi-dimensional feature adaptive network (RT-MDAFNet): Firstly, the Adaptive Fusion Pyramid Network (AFPN) is designed at the feature fusion layer of the model, and the adaptability and feature extraction ability of the model to multi-scale targets are improved through the dynamic channel attention mechanism and selective feature fusion mechanism. Then, the adaptive channel-spatial aggregation network module (SASE-RepNet) was designed to improve the detection accuracy and efficiency in complex backgrounds by combining multi-level feature aggregation, channel adaptive weight allocation and spatial selectivity enhancement mechanism. In the absence of existing datasets, a dataset of special-shaped electronic components was constructed, and the RT-MDAFNet model was compared with 8 baseline models, including DETR, Faster R-CNN, YOLO series, etc.