A lightweight surface defect detection framework combined with dual-domain attention mechanism

In this paper, a lightweight printed circuit board (PCB) defects detection model (light-PDD) is proposed, which mainly concentrates on overcoming the deficiencies of redundant parameters and slow inference speed in most existing methods. In particular, overall framework of the proposed light-PDD has followed the YOLOv4 model with further enhancement, and the backbone part adopts a pruned MobileNetV3 structure for feature extraction, where a dual-domain attention mechanism is designed and diverse activation functions are employed so that the model can effectively handle the difficulties in detecting tiny-size PCB defects. Moreover, the improved cross-stage partial structure is deployed for feature fusion, which removes redundant parameters and avoids duplication of gradient information so as to reduce the model complexity. Extensive experiments on public PCB defect datasets have demonstrated the superiority of the proposed light-PDD, which outperforms several other advanced algorithms in terms of realizing the balance between inference speed and detection accuracy. In addition, the size of the proposed light-PDD is only 95.7 MB, the inference speed is improved by 34.37 frames per second of light-PDD as compared to the baseline model YOLOv4 and simultaneously, a comparable accuracy of 97.13% is obtained, which verifies that the proposed light-PDD is a reliable and competitive lightweight model for PCB defects detection.

Automated summary

This paper proposes a lightweight printed circuit board defects detection model (light-PDD) to overcome the deficiencies of redundant parameters and slow inference speed in existing methods. The light-PDD model follows the overall framework of YOLOv4 with enhancements, using a pruned MobileNetV3 structure for feature extraction. It also incorporates a dual-domain attention mechanism and diverse activation functions to effectively handle the detection of tiny-size PCB defects. The improved cross-stage partial structure is deployed for feature fusion to reduce model complexity. Experimental results demonstrate the superiority of light-PDD in terms of inference speed and detection accuracy.