Abstract: Aiming at the problems of small size defects and warp and weft defects with extreme aspect ratios on the surface of towels,an image detection method for towel defects based on Convolutional Neural Network (CNN) was proposed.An area-scan industrial camera in the constructed towel defect detection system was used to collect towel images,and 150 images with defects were selected.After that,a dataset was made after data augmentation of the collected images.By fusing the feature pyramid network with ResNet-50 and introducing K-means clustering to optimize the aspect ratio of bounding box,the applicable Faster R-CNN target detection algorithm was obtained.To extract the defect features in images and identify the defect targets,the data set was put into network for training,and the trained network was experimentally verified,the recognition detection rate reached 95.2%.Experimental results showed that the system could effectively realize automatic detection of towel defects.

Key words: convolutional neural network, towel defect detection, Faster R-CNN, feature pyramid network, K-means clustering

摘要： 针对毛巾表面小尺寸瑕疵和极端纵横比的经纬向瑕疵,提出基于卷积神经网络的毛巾瑕疵图像检测方法。首先采用所构建毛巾瑕疵检测系统中的面阵工业相机进行毛巾样本图像采集,从中选取150张带有瑕疵的图像,对采集的图像进行数据扩充,并制作数据集;其次,通过融合特征金字塔网络与ResNet-50,并引入K-means聚类优化边界框宽高比,得到适用的Faster R-CNN 目标检测算法;最后,用数据集进行网络训练,提取图像中的瑕疵特征,识别瑕疵目标,并对训练所得网络进行实验验证,识别检出率达到95.2%。结果表明,所提出的系统可有效实现毛巾瑕疵自动检测。

关键词: 卷积神经网络, 毛巾瑕疵检测, Faster R-CNN, 特征金字塔网络, K-means聚类