基于 CR-RFPR101 的钢板表面缺陷检测

针对钢板表面缺陷种类多、背景复杂、检测精度低等问题, 文章首先对钢板表面缺陷数据集进行数据增强, 并对原始 Cascade 区域卷积神经网络 (region-based convolutional neural networks, R-CNN) 算法进行改进, 将 ResNeXt-101-64×4d 作为 Cascade R-CNN 算法的骨干网络, 优化特征提取模块, 利用递归特征金字塔 (recursive feature pyramid, RFP) 网络以反馈连接的方式进行特征优化, 提出一种 CR-RFPR101 (Cascade R-CNN RFP ResNeXt-101-64×4d) 的检测算法, 以更好地保留细节和语义信息; 同时使用可切换的空洞卷积替换主干网络的卷积层, 以改变感受野的方式提高检测性能; 最后使用引入软化非极大值抑制算法, 保留有效信息, 提高识别率。经实验验证, CR-RFPR101 算法的检测率为 83.4%, 比原 Cascade R-CNN 算法提高了 7.3%, 满足了钢板表面缺陷检测要求。

In view of the multiple types, complex background and low detection accuracy of steel plate surface defects, the data set is enhanced and the original Cascade region-based convolutional neural networks(R-CNN) algorithm is improved. ResNeXt-101-64×4d is used as the backbone network of Cascade R-CNN algorithm to optimize feature extraction module. The recursive feature pyramid(RFP) network is used to optimize the features in the way of feedback connection, and a detection algorithm based on Cascade R-CNN RFP ResNeXt-101-64×4d(CR-RFPR101) is proposed to better retain the details and semantic information. Switchable atrous convolution is proposed to replace the convolution layer of the backbone network to improve the detection performance by changing the receptive field. Finally, the soft non-maximum suppression(Soft-NMS) algorithm is introduced to retain the effective information and improve the recognition rate. The experiment shows that the detection rate of CR-RFPR101 algorithm is 83.4%, which is 7.3% higher than that of the original Cascade R-CNN algorithm. The proposed algorithm meets the requirements of steel plate surface defect detection.