基于破坏学习的残差网络丝饼毛羽缺陷分类  

DTY Package Hairiness Defect Classification Based on Destruction Learning of Residual Network

作　　者：张诗涵 景军锋[1] 宋智伟 ZHANG Shihan;JING Junfeng;SONG Zhiwei(Xi′an Polytechnic University,Xi′an,710600,China)

机构地区：[1]西安工程大学,陕西西安710600

出　　处：《棉纺织技术》2022年第3期8-13,共6页Cotton Textile Technology

基　　金：陕西省创新能力支撑计划(2021TD-29);陕西高校青年创新团队;大学生创新创业项目(201910709004)。

摘　　要：针对人工分类丝饼表面毛羽缺陷存在检测效率低、易漏检误检等问题,提出了一种基于破坏学习的残差网络丝饼毛羽缺陷分类方法。首先在网络的输入部分提出一种区域混乱机制,将输入图片划分为局部区域后在一定范围内随机打乱,以此更加凸显毛羽缺陷的局部细节特征;然后提出一种对抗损失函数消除由于区域混乱机制引入的噪声信息;最后使用全局最大池化代替残差网络的平均池化,加强对毛羽特征的提取能力,并利用softmax分类器进行分类。试验结果表明:本研究提出的方法在构建数据集下平均分类准确率达到95.0%,平均每张图片的测试时间为30 ms。认为:基于破坏学习的残差网络丝饼毛羽缺陷分类可以满足工业中的精度和实时性需求。Aiming at the problems of lower detection efficiency,easier to miss detection and false detection by manual method of DTY packages hairiness defect,a method for classification of hairiness defects in residual network based on destruction learning was proposed.Firstly,a regional confusion mechanism was proposed to divide the input image into local regions and randomly scramble within a certain range to inhance highlight local details of hairiness defects.Then,an adversarial loss function was proposed to eliminate the noise information introduced by the regional confusion mechanism.Finally,the global maximum pooling was used instead of the mean pooling of the residual network to strengthen the extraction ability of hairiness features,the softmax classifier was used for classification.The test results showed that the average classification accuracy of the proposed method was reached 95.0%under the constructed data set,the average test time for each image was 30 ms.It is considered that DTY packages hairiness defect classification based on destruction learning of residual network can meet the demand of accuracy and real-time in industry.

关 键 词：丝饼 毛羽分类 破坏学习 对抗损失函数 残差网络 

分 类 号：TP391.1[自动化与计算机技术—计算机应用技术]