基于轻量化卷积神经网络的毛巾织物瑕疵检测方法

Towel fabric defect detection method based on lightweight convolutional neural network

针对毛巾织物瑕疵检测中存在的小目标瑕疵漏检率高、形变尺度大的瑕疵检测精度低以及模型检测效率不理想等问题,提出一种基于YOLOv4网络的轻量化毛巾织物瑕疵检测方法。采用轻量级网络Ghost Net重构主干特征提取网络,以降低模型运算量,提升检测速度;在深层特征提取网络中引入结合空洞卷积和SoftP ool的DS-CBAM模块,扩大感受野的同时保证特征图分辨率并提高模型对毛巾织物瑕疵特征的提取能力;根据各类毛巾织物瑕疵正负样本不平衡的数据特点,引入难易样本聚焦参数和正负样本平衡参数对损失函数进行优化,降低样本失衡对检测性能的影响;采用改进度量距离的K-means算法自适应生成适合毛巾织物瑕疵尺寸的先验框,提高先验框和毛巾织物瑕疵目标的匹配度。研究结果表明:改进后的模型在毛巾织物瑕疵数据集上的检测精度要优于原YOLOv4和其他主流检测算法,综合类别平均精度达到92.14%,检测速度达到49.98帧/s,分别比原模型提高了5.31%、22.83%,有效平衡了检测精度和检测速度之间的关系。

In order to solve the towel fabric defect detection problems such as missed detection of small target defects,low detection accuracy of large deformation scale defects and unsatisfactory model detection efficiency,a lightweight towel fabric defect detection method based on YOLOv4 network is proposed.GhostNet,a lightweight network is used to reconstruct the backbone feature extraction network,to reduce the amount of model operations and improve the detection speed.A DS-CBAM modulecombined with cavity convolution and SoftPool is introduced into the deep feature extraction network to expand the perceptual field while ensuring the feature map resolution and improving the model's ability to extract towel fabric defect features;according to the data characteristics of positive and negative samples of various types of towel fabric defects.The K-means algorithm with improved metric distance is used to adaptively generate a priori frames suitable for towel fabric defect sizes and improve the matching between the a priori frames and towel fabric defect targets.The experimental results show that the improved model outperforms the original YOLOv4 and other mainstream detection algorithms in terms of detection accuracy on the towel fabric defect dataset,achieves combined category average accuracy of 92.14%and detection speed of 49.98 fps,which are 5.31%and 22.83%higher than the original model respectively,effectively balancing the relationship between detection accuracy and detection speed.