【作者】 张淑敏；

【导师】 高占恒；

【作者基本信息】 吉林大学 ， 计算机技术（专业学位）， 2020， 硕士

【摘要】 衣食住行对人们的日常生活至关重要,就衣物而言,布匹是制作衣物不可缺少的原材料。在布匹生产过程中,机械操作不当、客观环境等原因让纺织机在生产过程中使布匹表面出现疵点。布匹表面是否存在疵点是决定布匹质量好坏的一个非常关键的因素,但目前许多企业的检测方式还是人工检测,而且工人长期利用肉眼检测会产生视觉疲劳,以至于检测速度下降,不适合实际的生产。因此,寻找比较自动化的技术进行布匹检测显得尤为重要。布匹疵点检测方法大致可分为四大类,分别为统计方法、模型方法、频谱方法以及机器学习方法,但是统计方法、模型方法和频谱方法都存在计算量大以及实时性差等问题,不适合在实际生产中使用。随着现在计算技术的发展,尤其是深度学习的发展为布匹疵点检测提供了更新的、更可靠的技术支持。为了提高布匹疵点检测的性能,本文提出了一种对于特征金字塔进行改进的布匹疵点检测方法。针对布匹疵点检测中瑕疵目标尺度多变、形状不规则、存在较多的小目标等问题,本文将基于深度学习的目标检测算法应用到布匹疵点检测中。Faster R-CNN是基于深度学习的目标检测算法的经典算法之一,FPN是基于Faster R-CNN的改进,构建了多尺度的特征映射,形成特征金字塔结构,将包含较多分类语义信息的高层特征与包含较多位置信息、纹理颜色等信息的低层特征进行融合。实验表明,直接将FPN应用布匹疵点检测中会存在一定的问题,包括目标的漏检、错检等。针对这些问题,本文对FPN进行了一些改进,主要改进点为:针对数据集图像中疵点形状的不规则性,本文引入了可变形卷积,从而适应数据集中形状各不相同的疵点,改善了模型对不同形状瑕疵点的检测能力,增强了模型的泛化性能;同时针对单纯的特征金字塔结构,会存在特征层级的不平衡问题,会使得当前层的特征更加关注相邻层的语义信息,引入了Balanced Feature Pyramid模块,通过对不同级别的特征进行特征融合,再将融合后特征分配到各个层级,使得每层的特征都有来自各层的信息,从而提高检测性能。本文实验数据来自天池比赛布匹疵点检测的数据集,通过实验,Faster R-CNN和FPN在数据集上的mAP值分别为58.61和59.69,本文改进后的算法mAP值为64.30,比Faster R-CNN高了5.69个点,比FPN高了4.61个点,在小目标的检测中,本文改进的算法比FPN提高了2.1个点。实验结果表明,改进后的算法对布匹疵点检测有着良好的效果。更多还原

【Abstract】 Clothing,food and shelter are vital to people’s daily lives.As far as clothing is concerned,cloth is an indispensable raw material for making clothing.In the fabric production process,improper mechanical operation and objective environment cause the textile machine to cause defects on the surface of the fabric during the production process.Whether there is a defect on the surface of the fabric is a very critical factor that determines the quality of the fabric.However,the current inspection methods of many enterprises are still manual inspection,and workers’ long-term use of naked eyes will cause visual fatigue,so that the detection speed is reduced,which is not suitable for practical produce.Therefore,it is very important to find a more automated technology for cloth inspection.Fabric defect detection methods can be roughly divided into four categories,namely statistical methods,model methods,spectrum methods and machine learning methods.However,statistical methods,model methods and spectrum methods have problems such as large calculation volume and poor real-time performance.They are not suitable for use in actual production.With the development of current computing technology,especially the development of deep learning,it provides updated and more reliable technical support for fabric defect detection.In order to improve the performance of fabric defect detection,this paper proposes an improved fabric defect detection method for feature pyramids.Aiming at the problems of variable target size,irregular shape,and many small targets in cloth defect detection,this paper applies a object detection algorithm based on deep learning to fabric defect detection.Faster R-CNN is one of the classic algorithms of object detection algorithms based on deep learning.FPN is based on the improvement of Faster R-CNN.It builds multi-scale feature maps to form feature pyramid structures.It combines high-level features that contain more classified semantic information with low-level features that containing more location information,texture color,and other information.Experiments show that there will be certain problems in directly applying FPN to fabric defect detection,including missing targets,error detection,etc.In response to these problems,this article has made some improvements to FPN.The main improvement points such as : Aiming at the irregularity of the shape of the defect in the dataset image,a deformable convolution is introduced in this paper to adapt to defects with different shapes in the dataset,improve the model’s ability to detect defects of different shapes and enhance the model’s generalization performance;At the same time,for the simple feature pyramid structure,there will be an imbalance of feature levels,which will make the features of the current layer pay more attention to the semantic information of adjacent layers.This article introduces the Balanced Feature Pyramid module.This paper improves the detection performance by performing feature fusion on different levels of features,and then assigning the fused features to each level so that the features of each layer have information from each layer,thereby improving detection performance.The experimental data in this paper comes from the Tianchi competition cloth defect detection data set.Through experimental analysis,the mAP values of Faster R-CNN and FPN on the data set are 58.61 and 59.69,respectively.The improved algorithm’s mAP value is 64.30,which is 5.69 points higher than Faster R-CNN and 4.61 points higher than the FPN.In the detection of small targets,the mAP value of the improved algorithm in this paper is 2.1 points higher than the FPN.Experimental results show that the improved algorithm has a good effect on cloth defect detection.更多还原