基于生成对抗网络的高光织物表面疵点检测与识别

China is the world's largest textile exporter. In order to improve textiles quality and brand value, defects detection has become a very important part of the textile manufacturing process. In view of the severe interference of high gloss and texture, and the difficulty in the defects detection in high gloss fabrics and leather, this project aims to establish a new defects detection algorithm for textured and specular objects based on generative adversarial network. This project is proposed to solve the following key scientific problems: 1) Compared with the traditional highlight removal algorithm, it is a new idea to automatically remove the image highlight of specular objects through learning material rendering parameters. The problem is how to improve the existing image style migration algorithm to improve image conversion accuracy; 2) Fabric defects detection based on deep learning has just started. Without large-scale dataset, how to extract defects based on the small unpaired dataset becomes the key problem. By solving the above-mentioned questions, this project will not only establish a new system of defects detection to improve the defects recognition rate of the fabric and leather, but also can provide a new insight in highlight removal for the optical detection of specular objects. Meanwhile, a new fabric defects dataset, available for the relevant researchers, can be produced. Therefore, this project has important industrial application value and shows significant academic research value.

中国是世界最大纺织品出口国，为了提高纺织品质量以及品牌价值，疵点检测成为纺织品制造工序中非常重要的环节。针对目前高光织物、皮革的疵点检测面临的饱受高光、纹理干扰严重、检测难度大等问题，本项目拟基于生成对抗网络建立具有纹理和镜面反射特性物体的疵点检测新算法。拟解决关键科学问题为：1)相对于传统去高光算法，通过学习材质渲染参数自动去除物体图像的高光是一个全新思路，然而如何改进现有的图像风格迁移算法以提高图像转换精度；2)基于深度学习的织物疵点检测才刚刚起步，缺少大规模数据集，如何突破小样本非配对数据集提取织物瑕疵难的技术瓶颈。通过解决上述问题，本项目既可建立一套新的疵点检测系统以提高织物、皮革的疵点识别率，又可为镜面物体的光学检测提供新的去高光思路，还可建立一个新的织物疵点数据集供相关研究者使用，因此具有重要的工业应用价值和学术研究意义。

高光织物与皮革缺陷检测面临高光与纹理干扰严重等问题;本项目基于生成对抗网络建立了适合具有纹理和镜面反射特性物体缺陷检测问题的新模型。具体研究成果为：采集了一个真实场景下的高光泽皮革缺陷数据集；提出了基于改进生成对抗网络的图像风格迁移算法，建立了分别基于多视角序列图像和基于双掩码指导的两个高光去除网络；提出了一种基于特征增强的去偏融合网络以提高皮革缺陷识别率。本项目建立了一个新检测系统以提高高光织物与皮革的缺陷检出率，为镜面物体的光学检测提供了新的去高光思路，具有重要的工业应用价值和学术研究意义。