基于图像理解的织物缺陷智能检测识别方法研究

The intelligent detection and recognition of fabric defects are key factors of the quality control of textile products. Many existing methods are based on traditional image processing and pattern recognition methods, which suffer low detection and classification accuracies and poor adaptability. This project, meeting the industry demands with the latest developments in artificial intelligence community, is to deeply study the intelligent fabric defects detection methods, establish the fabric-defect-oriented deep learning model based on large sample set and human-like perception model based on small sample set. The focus of this project is to solve the key scientific problems of textile features of fabric images, adaptive deep learning model and reasoning model based on small sample set. On the basis of deep learning principles and for the problems of the multi-scales and diversities of fabric textures and defects, this research aims to establish a multiple-layer neural network model, which is robust to textile features, fabric defects, to solve the problem of fabric defects detection and recognition with large sample set. On the other hand, inspired by the human perception mechanisms, including human visual perception for image representation, defects detection via sparse representation and the intelligent fabric defects detection and recognition framework based on human-like-conception-learning, this research is to provide the new solutions to defects detection and recognition via small sample set. The research results mean to improve the intelligence level of fabric defects detection and contribute positively to the pattern recognition and computer vision community.

织物疵点智能检测识别是纺织品质量控制的关键环节，现有方法多采用传统图像处理及模式识别方法，检测识别的准确率低，且缺乏自适应性。本项目拟将人工智能领域的最新进展与行业需求相结合，深入研究织物疵点智能检测识别方法，建立面向织物疵点的大规模样本深度学习模型和小样本类人感知模型，重点解决织物图像纹理表征、深度模型自适应学习、小样本模型推理等关键科学问题。一方面，拟基于深度学习方法，针对实际生产过程中织物纹理和疵点多尺度、多样性问题，建立对纹理、疵点具有鲁棒性、且支持在线学习的多层次神经网络模型，解决大样本条件下织物疵点检测与识别问题；另一方面，拟借鉴人类感知机制，构建包括基于人类视觉感知的图像表征、稀疏表示的疵点检测、类人概念学习的疵点识别的智能检测识别框架，为基于小样本的疵点检测识别提供新的解决方案。相关研究成果有望提升织物疵点检测识别的智能化水平，并对模式识别与机器视觉领域产生积极影响。

复杂纹理织物图像中形态多样的缺陷检测是当前机器视觉领域的一个难题。本项目深入研究了面向织物缺陷的大规模样本深度学习模型和小样本类人感知模型，重点解决了深度模型自适应学习、小样本模型推理、模型轻量化等关键科学问题，取得了一系列创新性成果：1）基于公开和采集的大规模带标注织物缺陷图像集，提出基于特征迁移网络、深度可分离卷积神经网络及可分离自适应再校准模块的织物缺陷检测模型，提升了深度模型的检测精度和速度；2）基于深度学习技术，建模人类视觉显著性机制，通过多尺度判别力特征提取、双向信息交互、双分支平衡、多感受野特征细化等增强深度模型自适应学习能力，从而提升弱对比细小缺陷的检测效果；3）针对大规模带有像素级标签缺陷数据集难以构建的问题，基于多实例学习技术，提出三分支特征提取网络、自注意机制等特征增强策略的弱监督织物缺陷检测算法，仅依靠图像级标签，即可实现织物缺陷的快速分割定位；4）在提出检测模型的基础上，研究面向小样本模型推理的缺陷检测方法，基于元学习，提出了特征匹配模块、多层互补模块的小样本检测算法，提升小样本下的缺陷检测精度，解决样本采集难题；5）针对检测模型计算复杂和参数冗余，研究深度学习模型轻量化方法，提出基于特征图显著性、基于动态通道选择的滤波器剪枝方法，对检测模型进行压缩并加速，使其易于部署在资源受限的边缘设备上。课题研究成果最终突破复杂纹理织物缺陷检测在实际应用中的理论及技术瓶颈，提高了现有检测及分析方法的自适应性、检测精度和速度。围绕相关研究成果，共发表学术论文34篇，其中SCI检索18篇，EI检索15篇；申请国家发明专利8项，授权国家发明专利5项；获河南省科学技术进步一等奖、二等奖各1项，河南省自然科学学术论文二等奖1项，省优秀硕士论文1篇。基于提出的方法，开发了智能验布系统，已在河南、江苏、浙江和河北等地的纺织企业进行应用，产业化直接经济效益达到17857.0万元。