【作者】 魏巍；

【导师】 邓德祥；

【作者基本信息】 武汉大学 ， 电路与系统， 2021， 博士

【摘要】 棉花中的异性纤维(简称异纤)与织物上的疵点对棉纺织产品影响极大。在纺纱工序中异纤会被拉断或粉碎成无数细小的疵点,在织造阶段,这些疵点又会在棉纺织工业的最终产品—织物上形成肉眼难以分别的瑕疵,对棉纺织产品质量造成严重的影响。许多异纤与原棉本身及棉花杂质视觉特征接近,传统的人工特征设计方法误判严重,导致异纤无法精准有效清除;织物疵点种类繁多、样本收集困难,且同一类型疵点在不同的织物纹理上也有不同的特征区别,对检测算法设计的通用性和鲁棒性带来极大的挑战。因此,设计准确率高,通用性强且计算量适中的异纤与织物疵点检测模型具有重要的研究价值。本文主要针对多类异性纤维与织物疵点检测问题,主要工作总结如下:(1)在纺纱工序中,针对原棉中的异纤在图像特征上与棉花过于相似导致分类识别困难问题,提出了一种基于残差网络与特征差拟合的异纤分类模型。在类激活热力图映射算法基础上,通过卷积神经网络的特征可视化方法,结合异纤图像特点,观察经典网络对异纤图像的特征提取能力,对经典网络结构加以调整裁剪,设计了一种兼顾性能与效率的基础异纤分类网络Cotton Net,在验证集上的分类准确率达到94.2%。为了增强高层特征提取能力,本文在残差网络的特征融合方法基础上,改进了异纤分类网络并提出Cotton Net-Res,提升分类准确率到95.1%。最后针对复杂环境下的异纤分类问题,提出了一种基于特征差拟合的分类模型Cotton Net-Fusion,在普通异纤样本数据集中,分类准确率提升到97.4%,在复杂环境的异纤数据集上仍能保持90.3%的准确率。(2)针对传统原棉中异纤含量评估依赖人工抽样的局限性问题,提出了基于卷积神经网络的异纤图像分割算法。通过经典分割网络结构的比较与实验论证,提出了基于U-Net图像分割架构与目标位置回归网络的改进算法Cotton Res-YNet,利用图像中心点与目标相对位置的坐标回归思路,从坐标位置维度,进一步提升分割精度。解决了传统分割只对单像素分类从而造成轮廓信息丢失的缺陷,相对于传统的U-Net,分割指标均交并比m Io U(Mean Intersection over Union)从86.7%提高到90.3%。再根据分割后的异纤图像,结合分拣机机械与图像采集设备参数,计算每一个像素对应的物理尺寸,得到异纤的长度或面积,从而通过异纤的尺寸数据来评估原棉中异纤含量并进行统计分析,标定测试表明,算法测试结果与真实标定结果相对误差在4%以内,相对于传统的人工抽样方式,提出的算法可以更全面的对原棉异纤含量进行评估。(3)在织造工序中,针对织物种类过多,疵点数据收集困难等问题,提出了一种基于无监督学习的织物疵点检测算法。该算法通过无缺陷样本的自动学习训练,基于自动编码器的图像重构技术来突出疵点对织物的破坏。经典无监督网络中常使用像素距离误差损失函数来重构图像,但是这种方法无法反应织物缺陷对纹理结构的破坏。本文改进了经典无监督网络,引入结构相似性指标作为损失函数来训练网络重构图像的能力。实验表明,使用距离误差与结构相似性指标混合的重构评价函数在公开数据集以及本文数据集上性能最佳,验证数据集的检测准确率可达90%以上。(4)将异纤与织物检测相关研究集成到工业级的多核DSP平台系统上,提出了一种基于多核嵌入式平台的半静态调度算法,提升了多核并行效率,将核心网络Cotton Net-Res的计算时间优化在2ms内,满足了系统实时应用。以此为基础,实现了异纤分类设备并进行了真实异纤样本投放测试,实际异纤的分类准确率达到97.3%。通过异纤分割算法Cotton Res-YNet的实现,建立了基于机器视觉的异纤尺寸评估与国标数据之间的映射关系。在无监督学习的技术基础上,实现了织物疵点检测并进行了工业验证。综上所述,本文从棉花异性纤维检测—棉花质量评估—织物检测三个阶段针对棉纺织产品缺陷检测模型开展研究,最后实现了异纤和织物疵点检测系统,并完成测试和工业验证。在棉纺产品综合质量提升方面有重要的研究意义与应用价值。更多还原

【Abstract】 The foreign fibers in cotton and the defects on the fabric will seriously affect the quality of cotton textile products.In the spinning process,foreign fibers will be broken or shattered into numerous minor defects;in the weaving stage,these defects will form defects on the final product,difficult to distinguish by naked eyes,causing a severe impact on the quality of cotton textile products.Many foreign fibers are similar to the raw cotton itself and the visual characteristics of cotton impurities.Traditional artificial feature design methods often misjudge,which leads to the inaccurate and effective removal of foreign fibers.There are many kinds of fabric defects,which means it is difficult to collect samples.Even the same type of defects has different characteristics in different fabric textures,which brings significant challenges to the generality and robustness of detection algorithm design.Therefore,it is of great significance to design a defect detection model of foreign fiber and fabric with high accuracy,strong versatility,and a moderate amount of calculation.The main work is summarized as follows:(1)In the spinning process,a foreign fiber classification model based on residual network and feature difference fitting is proposed,aiming to solve the problem that the image features of foreign fibers in raw cotton are too similar to cotton,which leads to the difficulty of classification and recognition.On the basis of the class activation thermograph mapping algorithm,we design a primary foreign fiber classification network Cotton Net,via the characteristics of foreign fiber image and the feature visualization method of convolution neural network,and adjust and trim the classic network structure,and balance both performance and efficiency.The classification accuracy on the verification set reaches 94.2%.To enhance the ability of high-level feature extraction,we improve the foreign fiber classification network and propose Cotton Net-Res based on the feature fusion method of residual network.The classification accuracy is improved to 95.1%.Finally,aiming at foreign fiber classification in a complex environment,we propose a classification model based on feature difference fitting,Cotton Net-Fusion.In the ordinary foreign fiber sample data set,the classification accuracy is improved to 97.4%.In the foreign fiber data set of complex environments,it can still maintain the accuracy of 90.3%.(2)Aiming at the limitation of traditional evaluation of foreign fiber content in raw cotton relying on manual sampling,a foreign fiber image segmentation algorithm based on a convolution neural network is proposed.We propose an improved algorithm Cotton Res-YNet based on U-Net image segmentation architecture and target position regression network,through the comparison and experimental demonstration of classical segmentation network structure.Using the idea of coordinate regression of the relative position between the image center and target,the segmentation accuracy is further improved from the dimension of coordinate position.Compared with the traditional u-net,the segmentation index is improved from 86.7% to 90.3% compared with m Io U(Mean Intersection over Union).Then,according to the segmented foreign fiber image,combined with the sorter mechanical and image acquisition equipment parameters,we calculate the corresponding physical size of each pixel,and get the length or area of foreign fiber,so as to evaluate the content of foreign fiber in raw cotton through the size of foreign fiber and make statistical analysis.The calibration test shows that the relative error between the algorithm test results and the real calibration results is less than 4%.Compared with the traditional manual sampling method,the proposed algorithm can comprehensively evaluate the foreign fiber content of raw cotton.(3)In the weaving process,aiming at the problems of too many kinds of fabrics and the difficulty of collecting defect data,a fabric defect detection algorithm based on unsupervised learning is proposed.The algorithm highlights the damage of defects to the fabric through the automatic learning training of defect-free samples and the image reconstruction technology based on automatic encoder.In classical unsupervised networks,the pixel distance error loss function is often used to reconstruct images,but this method cannot reflect the damage of fabric defects to texture structure.We improve the classical unsupervised network,and introduce the structure similarity index as the loss function to train the network reconstruction image.The experimental results show that the performance of the reconstructed evaluation function based on distance error and structural similarity index is the best on both the public data set and the data set in this study,and the detection accuracy of the verified data set can reach more than 90%.(4)Integrating the research of foreign fiber and fabric detection into the industrial multi-core DSP platform system,we propose a semi-static scheduling algorithm based on multi-core platform,which improves the multi-core parallelism and optimizes the computing time of the core network Cotton Net-Res within 2ms,and meets the real-time requirement of the system.On this basis,the foreign fiber classification equipment is realized and the actual foreign fiber samples are put into the test.The actual foreign fiber classification accuracy reaches 97.3%.Through the implementation of the foreign fiber segmentation algorithm Cotton Res-YNet,the mapping relationship between the foreign fiber size evaluation based on machine vision and the national standard data is established.On the basis of unsupervised learning technology,fabric defect detection is realized and verified in industry.To sum up,we studied the defect detection model of cotton textile products in three stages,cotton foreign fiber detection,cotton quality evaluation and fabric detection,and finally realizes the foreign fiber and fabric defect detection system,and completes the test and industrial verification.It has essential research significance and application value in improving the comprehensive quality of cotton spinning products.更多还原