摘要
Abstract
To address the challenges of high missed detection rates and insufficient localization accuracy for small target defects on printed circuit board(PCB)surfaces,a detection model named MMNet is proposed based on Multi-Level Perceptual Atten-tion(MLPA)and Multi-Scale Feature Fusion(MSFF).This model aims to overcome performance bottlenecks caused by inad-equate multi-scale feature fusion,coarse-grained attention mechanisms,and limitations of loss functions in existing methods.The MLPA module is designed to enhance the model's focus on subtle defects through a synergistic mechanism combining local and global attention.Convolutional operations and global average pooling are utilized to extract multi-level features,and atten-tion weights are optimized through nonlinear transformations.The MSFF module is constructed to dynamically adjust the weights of multi-scale features using grouped convolutions and channel concatenation strategies,improving the robustness of feature rep-resentation.The Ratio-IoU loss function is proposed,and a ratio factor is introduced to optimize the width and height coverage range of the bounding box to improve the positioning accuracy.Experiments are conducted on the public dataset PKU-Market-PCB.The results show that MMNet achieves a detection accuracy of 97.28%,which is 4.49 percentage points higher than YOLOv8 and the mean Average Precision(mAP@0.5:0.95)reaches 68.14%,surpassing YOLOv8 by 4.73 percentage points.The infer-ence speed of the model is measured at 95 fps,meeting real-time industrial detection requirements.Ablation experiments are performed to validate the effectiveness of each module:MLPA improves mAP@0.5 by 3.87 percentage points,MSFF optimizes feature fusion efficiency,and the Ratio-IoU loss increases localization accuracy by 4.73 percentage points.By integrating multi-level attention mechanisms and dynamic feature fusion,the proposed method significantly enhances the detection performance for small target defects.It provides an efficient and reliable solution for the automated quality inspection of high-density PCBs.关键词
PCB缺陷检测/多层次感知注意力/多尺度特征融合/IoU损失函数Key words
PCB defect detection/multi-level perceptual attention/multi-scale feature fusion/IoU loss function分类
信息技术与安全科学
