具有共享和专有特征的纹理复杂的工业产品表面缺陷图像自适应识别研究

It is a challenge to improve the classification accuracy in surface defect images of industrial products, owing to the difficulty in segmenting the defects with complex textures. Motivated by the diverse categories of the complex textures, a novel adaptive method of segmentation on defects is proposed by a double low-rank matrix decomposition based on Laplacian regularizer. Based on the assumption that the shared features represent defect-free region and class-specific features represent defect region in surface defect images, an adaptive method of feature extraction on defects is developed by shared and class-specific dictionary learning, which is more discriminative than the traditional dictionary learning. As its discriminative ability of feature vectors is further enhanced by the sparse autoencoder, the accurate and efficient classification is then obtained by the artificial neural network. The research above will provide theoretical basis and technological support for the method development of surface defect recognition for the surface defect image with complex texture that has both shared and class-specific feature, while it has important academic significance and application value.

本项目针对当前表面缺陷图像识别方法对纹理复杂的表面缺陷图像中缺陷目标难以分割而造成分类精度不理想的问题而进行。由于表面缺陷图像具有复杂纹理以及缺陷形态的多样化，本项目提出基于拉普拉斯约束的双低秩矩阵分解模型来实现对缺陷目标的自适应分割。考虑到表面缺陷图像具有表征正常区域的共享特征和表征缺陷区域的专有特征，本项目提出基于共享和专有字典学习的稀疏表示模型来实现对缺陷目标的自适应特征提取，这样处理比传统的字典学习具有更好的类别区分度。将得到的特征矢量进一步采用稀疏自编码器来增强其类别判别力，根据人工神经网络来实现对缺陷目标的准确高效分类。本项目可以为具有共享和专有特征的纹理复杂的工业产品表面缺陷图像自适应识别方法的开发提供理论基础和技术支撑，具有重要的学术意义和应用价值。

本项目针对当前纹理复杂的表面缺陷图像中缺陷目标需要精准识别的需求，借鉴图像分解的思想，利用表面缺陷图像具有表征正常区域的共享特征和表征缺陷区域的专有特征的特点，从理论、算法、应用这三个层面系统研究基于矩阵分解的表面缺陷图像识别方法，为工业产品表面质量在线检测提供新思路和新技术。.由于表面缺陷图像具有复杂纹理以及缺陷形态的多样化，本项目提出了一种基于双低秩矩阵分解的钢板表面缺陷图像分割方法，实现了表面缺陷图像的准确分割；同时，研究了一种基于共享和专有字典学习的钢板表面缺陷图像分类方法，解决了表面缺陷图像的高效分类；另外，构建了一种钢板表面缺陷图像自适应识别方法，提高了表面缺陷图像的检测精度和效率。.依托本项目，协助培养博士研究生2人、硕士研究生3人；申请中国专利8项；发表SCI论文5篇；以副主编身份出版专著1部。