【摘要】 针对印刷电路板缺陷尺寸微小、形态复杂多样和低区分度导致检测准确率低、漏检率高等问题，提出了一种基于大核分离和通道先验卷积注意的印刷电路板缺陷检测方法。首先，结合多尺度特征提取和空间卷积注意力机制，提出大核分离空间金字塔池化以提升模型的多尺度特征整合能力和建模能力。其次，在Neck网络中构建P2小目标检测层，使模型学习更加丰富且鲁棒的特征表示。引入通道先验卷积注意力模块在通道和空间维度上动态分布注意力权重，保留通道先验的同时有效地提取空间关系，提高模型对小目标缺陷的检测精度。实验结果表明，本文方法在PKU-Market-PCB数据集上的mAP达到了98.6%,比基准模型YOLOv8n提升了3.4%,精确度提升了2.6%,召回率提升了4.6%,单张图像推理时间仅为4.1 ms,适于实时检测。该方法显著提高了印刷电路板缺陷检测的准确率和实时性，具有较高的实际应用价值。更多还原

【Abstract】 Addressing the issues of small defect size, complex form, and low discriminability in printed circuit boards that lead to low detection accuracy and high false positive rates, a PCB defect detection method based on large kernel separation and channel prior convolutional attention is proposed. First, combining multi-scale feature extraction and spatial convolution attention mechanism, large kernel separation spatial pyramid pooling is proposed to enhance the multi-scale feature integration ability and modeling capability of the model. Second, the P2 small object detection layer is constructed in the neck network to enable the model to learn richer and more robust feature representations. The introduction of channel prior convolutional attention modules dynamically distributes attention weights along both the channel and spatial dimensions, retaining channel prior information while effectively extracting spatial relationships, thereby enhancing the detection accuracy of small object defects in the model. The experimental results indicate that the mAP of the proposed method on the PKU-Market-PCB dataset reached 98.6%, outperforming the baseline model YOLOv8n by 3.4%. The precision is improved by 2.6%, and the recall is increased by 4.6%. The inference time per image is only 4.1 ms, making it suitable for real-time detection. In summary, this method significantly enhances the accuracy and real-time performance of printed circuit board defect detection, providing high practical application value.更多还原