基于改进Cascade Faster R-CNN的铝型材表面缺陷识别研究

Research on Surface Defect Recognition of Aluminum Profile Based on Improved Cascade Faster R-CNN

在铝型材实际生产过程中,由于碰撞、加工温度、压力等原因,可能导致铝型材产生擦花、脏点、喷流等数种表面缺陷,缺陷目标较小,长宽大,传统目标检测算法的准确率较低,严重影响铝型材的美观和质量。在Faster R-CNN网络的基础上,引入了多阶段模型训练方法使部分无缺陷样本生成对抗样本,用ResNeXt105网络代替原始VGG16网络提取图像特征,设计了Cascade Faster R-CNN的网络结构,采用FPN提取多尺度特征图并进行特征图融合。实验结果表明,在2722张图像测试集上,Faster R-CNN模型准确率为62.7%,网络模型测试准确率达到81.4%,提高了18.7%。故相比于其他网络模型,改进后的Cascade Faster R-CNN的模型具有更强的特征提取能力和泛化能力,为类似小目标检测提高了技术参考。

In the actual production process of aluminum profiles,due to collision,processing temperature,pressure and other reasons,aluminum profiles may produce several kinds of surface defects,such as scratch,dirty spots and jet.The defect target is small,long and wide,and the accuracy of traditional target detection algorithm is low,which seriously affects the appearance and quality of aluminum profiles.Based on the faster R-CNN network,a multi-stage model training method was introduced to make some defect free samples generate confrontation samples.ResNeXt105 network was used to replace the original VGG16 network to extract image features.The network structure of Cascade Faster R-CNN was designed.The multi-scale feature map was extracted by FPN and the feature map was fused.In 2722 image test sets,the accuracy of fast R-CNN model is 62.7%,and the accuracy of this network model is 81.4%,which is improved by 18.7%.Compared with other network models,the improved Cascade Fast R-CNN model has stronger feature extraction ability and generalization ability,which improves the technical reference for similar small target detection.