基于深度神经网络的素色布匹瑕疵检测算法研究

目的为了改进当前布匹检测算法覆盖瑕疵种类不全、瑕疵检测准确率低和定位精度差的问题,提出一种端到端的素色布匹瑕疵检测的实用算法。方法首先通过图像增强扩充样本数量,使用以Resnet50为主干的Cascade-RCNN网络,通过加入可变形卷积、特征融合网络,增加锚框数目的方法实现素色布匹瑕疵检测。结果通过实验对比表明,该算法可实现检测20种布匹瑕疵,检测是否为瑕疵布匹的准确率为97%,瑕疵定位的平均检测精度为65%,每张样本平均时间为80 ms。结论该算法有效提升了布匹瑕疵检测的准确率和精度,检测瑕疵类别更全面,并且可以获取缺陷位置和类别,能够满足工业上的生产需求。

Defect Detection Algorithm of Plain Cloth Based on Deep Neural Network

The work aims to propose an end-to-end practical algorithm for plain cloth defect detection in order to solve the problems in the current cloth detection algorithm including incomplete coverage of defect types, low defect detection accuracy and poor positioning accuracy. Firstly, the number of samples was expanded by image enhancement, the Cascade-RCNN network with Resnet50 as the backbone was used, and the method of adding deformable convolution and feature fusion network to increase the number of anchor frames realized the defect detection of plain cloth. The experimental comparison showed that the algorithm could detect 20 kinds of cloth defects, the accuracy of detecting whether the cloth was defective was 97%, the average detection accuracy of defect location was 65%, and the average time for detecting each sample was 80 ms. This algorithm effectively improves the accuracy and precision of cloth defect detection and detects more defect categories, and can obtain defect locations and categories, thus meeting industrial production needs.