基于MF⁃SSD网络的织物疵点检测  

Fabric Defect Detection Based on MF-SSD Network

作　　者：黄汉林 景军锋[1] 张缓缓[1] 王震 周君 HUANG Hanlin;JING Junfeng;ZHANG Huanhuan;WANG Zhen;ZHOU Jun(Xi′an Polytechnic University,Shaanxi Xi′an,710600)

机构地区：[1]西安工程大学,陕西西安710600

出　　处：《棉纺织技术》2020年第12期11-16,共6页Cotton Textile Technology

基　　金：国家自然科学基金项目(61902302);陕西高校青年创新团队;陕西省教育厅服务地方专项计划项目(19JC019)。

摘　　要：为解决实际工业环境纺织品表面疵点检测速度慢、漏检率高的问题,利用MobileNet的深度可分离卷积取代传统SSD主干网络(VGG⁃16)中的普通卷积加速提取织物疵点特征,对MobileNet主干网络中的不同尺度的卷积特征层进行上采样,获得融合特征图并通过下采样构建特征图像金字塔网络,进而提取不同尺度特征;最后,使用具有不同尺度的特征层对大小不同的疵点做预测。将该算法分别在公共数据集和个人构建数据集进行测试,并与目前主流目标检测算法进行对比。结果表明:对于输入尺寸为300 pixel×300 pixel的织物图像,该MF⁃SSD网络的平均准确率均值达到90.1%,单张检测时间为30 ms。认为:MF⁃SSD网络具有更快的检测速度和较高的准确率。In order to solve the problems like lower detecting speed and higher omission ratio of textiles surface defect detection in practical industrial environment,the deep separable convolution in MobileNet was used to replace the normal convolution in traditional SSD backbone network(VGG-16)to speed up the extraction of fabric defect characteristics.Up sampling was performed on the convolution characteristics with different scales in the backbone network of MobileNet.Fuse characteristic patterns were obtained and the feature image pyramid network was established by down sampling.Then different scale characteristics were extracted.In the end,the characteristic layer with different scales were used to predict different size of defects.The algorithm was respectively used in public dataset and individual dataset for testing.And it was compared with current mainstream object detection algorithm.The test results showed that the mean average precision of the MF-SSD network was reached up to 90.1%for the fabric images with the input size of 300 pixel×300 pixel.The detection time for single image was 30 ms.It is considered that MF-SSD network detection has higher detection speed and higher accuracy rate.

关 键 词：疵点检测 SSD网络 MobileNet 卷积神经网络 融合特征 织物检测 

分 类 号：TS101[轻工技术与工程—纺织工程] TP391.1[轻工技术与工程—纺织科学与工程]