MSA-DETR: multi-scale alignment detection transformer for PCB defect detection with enhanced cross-scale feature fusion

Abstract

Abstract Printed circuit board (PCB) surface defect detection is critical for ensuring product quality. PCB defects are challenging to detect due to their small size, sparse features, and high similarity to background textures. Existing deep learning approaches suffer from limited feature extraction, insufficient multi-scale fusion, and inadequate spatial alignment for small defects. To address these challenges, this paper proposes multi-scale alignment detection Transformer (MSA-DETR), a multi-scale alignment object detection framework. First, an adaptive multi-scale residual (AMSR) block is designed for the backbone network, which enhances cross-regional modeling capability and preserves high-frequency detail information through large selective kernel mechanisms and multi-level residual connections. Second, a small-object alignment and frequency-enhanced pyramid is constructed to achieve effective multi-scale feature alignment through spatial feature redistribution and cross-scale multi-directional receptive field fusion. Finally, a context-aware spatial calibration mechanism is designed to realize collaborative calibration of semantics and geometry via global context retrieval and learnable deformation alignment. Experimental results on two benchmarking datasets demonstrate significant improvements. On DsPCBSD+, MSA-DETR achieves 86.1% mAP0.5 and 52.4% mAP0.5:0.95, improving by 2.6% and 2.7% respectively over RT-DETR baseline, while on HRIPCB it reaches 98.4% mAP0.5 and 55.8% mAP0.5:0.95, improving by 2.3% and 1.8% respectively. The model uses 7.0% fewer parameters than the baseline. These results validate the effectiveness of the multi-scale alignment strategy in improving small object detection accuracy, providing a robust solution for industrial PCB microscopic defect inspection.