Fabric defect detection based on faster RCNN

Abstract

Considering that the traditional detection of fabric defect can be time-consuming and less-efficient, a modified faster regional-based convolutional network method (Faster RCNN) based on the VGG structure is proposed. In the paper, we improved the Faster RCNN to suit our fabric defect dataset. In order to reduce the influence of input data for Faster RCNN, we expanded the fabric defect data. Meanwhile, by taking the characteristics of the fabric defect images, we reduce the number of anchors in the Faster RCNN. In the process of training the network, VGG16 can extract the feature map through the 13 conv layers in which the activation function is Relu, and four pooling layers. Then, the region proposal network (RPN) generates the foreground anchors and bounding box regression, and then calculates the proposals. Finally, the ROI Pooling layer uses the proposals from the feature maps to extract the proposal feature into the subsequent full connection and softmax network for classification. The experimental results show the capability of fabric defect detection via the modified Faster RCNN model and indicate its effectiveness.