An Improved YOLOv11 Model with Dual-Backbone Architecture and SimAM Attention for Adaptive PCB Defect Detection

Abstract

Printed Circuit Board (PCB) defect detection is a critical task in electronics manufacturing quality control. However, existing detection methods face significant challenges in identifying extremely small defects (below 0.5 mm) and balancing detection accuracy with real-time performance. This paper proposes an improved YOLOv11-based model that integrates a dual-backbone architecture, a dedicated small-defect detection layer, and a parameter-free attention mechanism for adaptive PCB defect detection. Specifically, we design a "texture stream + semantic stream" parallel dual-backbone architecture, combining CSPDarknet-Tiny for local texture extraction and MobileNetV4-Small for global circuit context. A P2 detection layer with a stride of 4 is introduced to enhance the detection of sub-0.5 mm defects. Furthermore, the SimAM (Simple, Parameter-Free Attention Module) is embedded at the fusion points to refine feature representations without additional parameters. An adaptive late-stage fusion strategy with learnable Sigmoid-normalized weights enables end-to-end optimization. Extensive experiments on the DeepPCB dataset demonstrate that the proposed model achieves 91.6% mAP50, 59.8% mAP50-95, and 68 FPS with only 4.1M parameters and 8.7 GFLOPs, outperforming the baseline YOLOv11-n by 3.4% mAP50 and 4.0% mAP50-95, and surpassing recent state-of-the-art methods including HDH-YOLO and lightweight YOLO11 variants.