Dynamic Multi-dimensional Hybrid Attention for High-Precision PCB Defect Detection

This study looks into ways to handle multi-scale feature blurring, edge detail loss, and inaccurate localization in PCB defect detection. We introduce a Dynamic Multi-dimensional Hybrid Attention (DMHA) mechanism and build the DMHA-YOLO network based on the YOLOv12 framework. The DMHA mechanism mainly consists of three components: first, a hierarchical adaptive dynamic channel compression module that optimizes the expression of multi-scale defect features; second, an edge-enhanced spatial perception module with learnable weights that helps reduce detail loss; and third, a coordinate positional attention module with explicit row-column encoding that improves the localization of micro-defects. These components are put together in a cascade-residual design to allow interaction across channel, spatial, and positional dimensions. Evaluations on the Peking University PCB_DATASET reveal that DMHA-YOLO attains a 3% enhancement (from 91.1% to 94.1%) in detection accuracy and a 3.9% improvement (from 50.4% to 54.3% ) in mAP@50-95 compared to the baseline YOLOv12. When compared with modified models that use attention mechanisms like CBAM, CoordAtt, and ECA in the same framework, DMHA-YOLO achieves higher recall and better localization at different scales without increasing computation. Our experiments show that the method is effective and stable for PCB defect detection, and it a good choice for accurate quality checks in factories.