基于Swin Transformer和多支路空洞卷积的工业表面缺陷检测方法

Industrial Surface Defect Detection Method based on Multi Branch Dilated Convolution and Swin Transformer

关键词:  目标检测  缺陷检测  特征融合  swin transformer  Faster R-CNN

Key words:  object detection  defect detection  feature fusion  swin transformer  Faster R-CNN

摘要:

针对工业表面缺陷中缺陷尺寸差异大,种类多,容易出现漏检和准确率低等问题,提出了一种基于Swin Transformer和多支路空洞卷积的工业表面缺陷目标检测(SAMB-IOD)方法。使用滑动窗口自注意力机制和空洞卷积搭建骨干特征提取网络,以更好的获得全局上下文的信息。在颈部搭建多支路空洞卷积特征金字塔网络进行特征融合,旨在获得融合低层次的高分辨率特征和高层次丰富语义信息。采用感兴趣区域对齐(ROI Align)替代原始的感兴趣区域池化(ROI Pooling),减少量化操作的误差,以提升检测精度。优化损失函数,减少分布不平衡对模型性能的影响。使用GC10-DET和DeepPCB工业数据集验证改进算法的有效性。实验结果表明,在GC10-DET数据集和DeepPCB数据集上,mAP分别提升了15.7%和5.7%。该方法能够有效检测工业表面缺陷,满足工业制造复杂环境下的检测需求。

Abstract:

To address the challenges in industrial surface defect detection, such as large variations in defect size, numerous types of defects, missed detections, and low accuracy, an industrial surface defect detection method based on multi branch dilated convolution and Swin Transformer was proposed, namely SAMB-IOD (Industrial surface defect object detection network based on self attention and multi branch dilated convolution, SAMB-IOD). The backbone feature extraction network is built using the Shifted Window Transformer Block and dilated convolution, which better captures global contextual information. The multi branch dilated convolutional feature pyramid network is employed in the neck to achieve feature fusion, combining high-resolution low-level features with rich semantic information from high-level features. Region of interest alignment (ROI Align) is adopted to replace the original region of interest pooling (ROI Pooling) to reduce quantization errors and improve detection accuracy. Additionally, Optimize the loss function to reduce the impact of imbalanced distribution on model performance. The proposed algorithm"s effectiveness is validated on the GC10-DET and DeepPCB industrial datasets. Experimental results show that the mAP is improved by 15.7% and 5.7% on the GC10-DET and DeepPCB datasets, respectively. This method effectively detects industrial surface defects, meeting the detection requirements of complex manufacturing environments.