应用于产品表面缺陷检测的神经网络IBS-Net

Neural Network IBS-Net Applied to Product Surface Defect Detection

通过将深度学习的两阶段目标检测算法应用于表面缺陷检测中,并依据产品表面缺陷的特性改进网络,提出了IBS-Net算法,实现缺陷的分类识别与定位。IBS-Net改进在于提出了特征相关的非极大抑制方法(FR-NMS)和正样本扩充方法(PSA),依赖特征层间语义关系筛选候选框,将含有局部缺陷信息的候选框作为半正样本以辅助分类任务,体现由部分缺陷推知整体缺陷的思路;其次,利用缺陷之间的互斥性,提出了多类别非极大抑制方法(CR-NMS)应用于后处理阶段,以优化预测结果;此外,利用缺陷之间的重要性差异,改进了表面缺陷检测评估方法。实验结果表明:IBS-Net对13类芯片表面缺陷和6类热轧钢带表面缺陷的检测综合精准度分别达94.8%和89.2%,证明本算法具有良好的有效性和工程应用价值。

By applying the two-stage object detection algorithm of deep learning to the surface defect detection task, and improving the two-stage network according to the characteristics of product surface defects, the IBS-Net algorithm was proposed to realize the classification and location of defects.IBS-Net improvement included the feature-related non-maximum suppression(FR-NMS) and the positive sample expansion method, which rely on the semantic relationship between feature layers to screen proposals, and use the proposals containing local defect as semi-positive samples to assist the classification task, reflecting the idea of inferring the overall defect from partial defects.And by using the mutual exclusion between defects, a multi-class non-maximum suppression method(CR-NMS) was proposed to be applied in the post-processing stage to optimize the detection results.In addition, the surface defect detection evaluation method was improved according to the difference in importance between defects.Experimental results show that IBS-Net can detect 13 types of chip surface defects and 6 types of hot-rolled steel strip surface defects with a comprehensive accuracy of 94.8% and 89.2%,respectively, which demonstrates that it has good effectiveness and engineering application value.