基于胶囊网络与滑动窗口的织物疵点检测方法

Fabric Defect Detection Method Based on Capsule Network and Sliding Window

针对卷积神经网络因池化过程造成提取特征不充分,导致织物疵点检测精度差的问题,提出一种基于胶囊网络与滑动窗口的织物疵点检测方法。应用K-means聚类方法对标注的目标框进行聚类分析,得到合适的滑窗切分尺寸;应用滑动窗口按照一定尺寸切分疵点图像,以此实现后续的疵点定位;将得到的不同尺寸疵点图像输入胶囊网络实现疵点分类,并应用NMS算法去除冗余框得到疵点检测结果。基于Pytorch框架构建胶囊网络模型,并就各种模型参数对织物疵点分类准确率的影响进行分析,利用在实际场景中采集的各类疵点图像,对胶囊网络织物疵点检测模型进行验证,得到4种疵点的检测结果。为了证明该方法的泛化性,在TILDA数据集上进行验证,并与目前主流目标检测模型的检测精度进行对比。结果表明:该方法能够准确、有效地对织物疵点进行检测和分类识别,平均精度均值达到90.78%,比大部分主流目标检测模型的检测精度高。

Aiming at the problem of inadequate feature extraction and low defect detection accuracy caused by pooling process of convolutional neural network,a fabric defect detection method based on capsule network and sliding window was proposed.Firstly,the K-means clustering method was used to cluster the marked target frame,and the appropriate sliding window segmentation size was obtained.Secondly,the sliding window was used to slice the defect image according to a certain size,so as to realize the subsequent defect location.Finally,the defects images of different sizes were input into the capsule network to realize the classification of defects,while the NMS algorithm was used to remove the redundant frames and obtain the defect detection results.The capsule network model was built based on Pytorch framework,while the influence of various model parameters on the accuracy of fabric defect classification was analyzed.The capsule network fabric defect detection model was verified by using various defect images collected in the actual scene and the detection results of four kinds of defects were obtained.In addition,the detection accuracy of the proposed method was compared with that of the current mainstream target detection models.The results showed that the proposed method could detect and classify fabric defects accurately and effectively,with an average accuracy of 90.78%,which was higher than that of most mainstream target detection models.