SDB-YOLO: A Lightweight X-Ray Image Component Detection Algorithm Based on Semantic Dual-Branch Features

With the widespread application of advanced packaging technologies such as BGA in aerospace electronic equipment, the demand for automatic detection of industrial components in X-ray images has become increasingly prominent. This paper addresses the characteristics of such components, which exhibit regular geometric contours and simple grayscale distributions, as well as challenges like noise interference in industrial scenarios. We propose a semantic dual-branch feature target detection algorithm named SDB-YOLO (Semantic Dual-Branch-YOLO), capable of simultaneously extracting grayscale and contour features. The dual-branch feature extraction network structure consists of the Grayscale Feature Branch (GFB) for grayscale feature extraction and the Contour Feature Branch (CFB) for geometric contour feature extraction. In the Grayscale Feature Branch (GFB), the network architecture is optimized through a structured pruning method, constructing a dedicated grayscale feature extraction network that reduces model parameters while improving detection accuracy. The Contour Feature Branch (CFB) incorporates a parallel wavelet transform module for geometric contour feature extraction and deeply fuses these features with grayscale features, significantly enhancing the model's feature representation capability. Comparative experiments on a custom dataset demonstrate that the improved algorithm achieves an average precision ([email protected]:.95) of 79.4%, representing a 2.1% improvement over the original YOLOv5. In small-sample scenarios, the [email protected](%) increases by 54.1%, while the model parameter count is reduced by 78.9%.