基于改进的Faster RCNN碳纤维编织物缺陷检测

Defect inspection of carbon fiber braided fabric based on improved Faster RCNN

以碳纤维立体编织物表面质量检测为研究对象,针对传统视觉识别率不高、小缺陷特征定位和识别不够准确的问题,提出基于改进的Faster RCNN缺陷检测算法。选取ResNet50作为主干特征提取网络,解决小缺陷特征图在卷积操作中的失真情况;采用Soft-NMS算法替换传统的NMS算法,以降低缺陷重复图像的误删;采用RoI Align层替换RoI Pooling层,以降低两次量化取整操作带来的缺陷特征定位误差。在PyTorch框架上对PASCAL VOC2007格式的缺陷图像数据集进行训练和测试。结果表明:改进的Faster RCNN平均精度均值为92.7%,相比原始Faster RCNN网络提升了3.8个百分点,其中毛刺等小缺陷特征识别平均精度提升了5.6个百分点。认为:基于改进的Faster RCNN模型可以满足碳纤维立体编织物表面质量检测要求。

Surface quality detection of carbon fiber braided fabric was taken as research object.Aiming at the problems of lower recognition rate of traditional vision,inaccurate positioning and recognition of smaller defect features,a defect detection algorithm based on improved Faster RCNN was proposed.ResNet50 was selected as the main feature extraction network to solve the distortion of smaller defect feature map in convolution operation.SoftNMS algorithm was used to replace traditional NMS algorithm to reduce the false deletion of repeated defect images.RoI Pooling layer was replaced by RoI Align layer to reduce defect feature positioning error caused by two quantization operations.The PASCAL VOC2007 format defect image data set was trained and tested on PyTorch framework.Results showed that the mean average precision of the improved Faster RCNN is 92.7%,which is 3.8percentage points higher than that of original network,and the mean average precision of small defect features such as burrs was increased by 5.6 percentage points.It is considered surface quality detection requirements of carbon fiber braided fabric can be met based on improved Faster RCNN model.