【作者】 王超；

【导师】 顾正晖；

【作者基本信息】 华南理工大学 ， 电子信息（专业学位）， 2023， 硕士

【摘要】 产品质量检测是现代纺织品生产中的重要环节,相比于人工质检法和传统机器视觉方法,基于深度学习的布料瑕疵检测算法因其准确率高、适用性强等优势,逐渐成为行业研究热点。而实际生产中正负样本不均衡、缺少标注信息等问题为高质量的布料瑕疵检测带来不小挑战。近年来,预训练模型因其丰富的知识和强大的泛化能力在许多任务中大展拳脚。本文将基于预训练模型的蒸馏学习方法应用到布料瑕疵检测任务中,根据数据集特点和任务需求设计基于知识蒸馏的布料瑕疵检测和定位算法,无需真实异常样本训练即可达到稳定高效的检测效果。具体研究内容包括:(1)针对布料数据正负样本不均衡的问题以及系统对瑕疵检测精度和速度的高要求,本文提出一种基于复习知识蒸馏的布料瑕疵检测与定位算法。算法利用预训练教师模型和学生模型中低层输出特征间的差异实现了布料瑕疵检测与定位。为充分保证各级正常样本有效知识的传递,提出了一种基于注意力融合的跨级知识复习模块对学生网络特征进行改进,并针对该模块设计了一种层级上下文复习训练约束。实验结果表明在私有涤纶和棉布数据集上,所提算法相比现有的代表性无监督瑕疵检测方法,在保证较高的检测效率的同时,能够取得更加优异的瑕疵检测性能。(2)针对现有无监督瑕疵检测算法在相对难检的灰布布料上表现不佳的问题,本文提出了一种基于柏林噪声算法的异常模拟策略,通过构造伪瑕疵开展训练以突破现有无监督学习算法的性能瓶颈。所提策略通过破坏正常布料的纹理和结构生成异常数据源,再结合柏林噪声掩模与正常布料图像软融合得到伪瑕疵样本。方法有效拓宽了布料瑕疵检测任务的设计和解决思路。(3)基于前述知识蒸馏框架和异常模拟策略,本文提出了一种全新的自监督布料瑕疵分割算法。为充分利用师生网络各级特征的差异,算法设计了横向连接模块对原始多级差分特征进行注意力编码和特征融合。此外,算法从充分保留图像细节信息以及缓解瑕疵前后景不均衡问题的角度设计训练约束。最终实现了端到端的布料图像瑕疵检测与分割。实验结果验证了所提异常模拟策略和算法设计的有效性,能够显著提升现有算法在相对难检的灰布布料上的瑕疵检测与定位性能。更多还原

【Abstract】 Product quality inspection is an important part of modern textile production.Compared to manual inspection methods and traditional machine vision methods,deep learning-based fabric defect detection algorithms have become a hot topic in the industry due to their high accuracy and strong applicability.However,the imbalance of positive and negative samples and the lack of labeling information in actual production pose significant challenges to highquality fabric defect detection.In recent years,pre-trained models have demonstrated their rich knowledge and powerful generalization capabilities in many tasks.This dissertation applies the distillation learning method based on pre-trained models to the task of fabric defect detection,and designs a fabric defect detection and localization algorithm based on knowledge distillation according to the characteristics of the dataset and task requirements.It can achieve stable and efficient detection results without training on real abnormal samples.The specific research content includes:(1)In view of the problem of imbalanced positive and negative samples in fabric data and the high requirements for defect detection accuracy and speed,this dissertation proposes a fabric defect detection and localization algorithm based on review knowledge distillation.The algorithm uses the differences between the low-level output features of pre-trained teacher and student models to achieve fabric defect detection and localization.In order to fully ensure the transmission of effective knowledge of normal samples at all levels,a cross-stage knowledge review module based on attention fusion is proposed to improve the features of the student network,and a hierarchical context review training constraint is designed for this module.Experiments on private polyester and cotton datasets have proved that compared with existing representative unsupervised defect detection methods,the proposed algorithm can achieve better defect detection performance while ensuring high detection efficiency.(2)To address the problem that existing unsupervised defect detection algorithms do not perform well on relatively difficult to detect fabrics such as grey cloth,this dissertation proposes an anomaly simulation strategy based on the Berlin Noise algorithm,which is trained by constructing pseudo-defects to break through the performance bottleneck of existing unsupervised learning algorithms.The proposed strategy generates abnormal data sources by disrupting the texture and structure of normal fabrics,and then obtains pseudodefect samples by soft fusion of Perlin noise masks and normal fabric images.The method effectively broadens the design and solution ideas for fabric defect detection tasks.(3)Based on the aforementioned knowledge distillation framework and anomaly simulation strategy,this dissertation proposes a novel self-supervised fabric defect segmentation algorithm.In order to fully utilize the differences in the features of the teacherstudent network at all levels,the algorithm designs a lateral connection module to perform attention encoding and feature fusion on the original multi-level difference features.In addition,the algorithm designs training constraints from the perspective of fully retaining image detail information and alleviating the imbalance between foreground and background of defects.Finally,end-to-end fabric image defect detection and segmentation is achieved.Experiments have verified that the effectiveness of the proposed anomaly simulation strategy and algorithm design,and can significantly improve the performance of existing algorithms in detecting and locating defects on relatively difficult-to-detect gray cloth fabrics.更多还原