基于ContrastGAN的色织物缺陷检测

Defect Detection of Yarn-dyed Fabric Based on ContrastGAN

传统自编码器应用于图案、背景纹理复杂的色织物缺陷检测任务中,存在普适性差以及漏检率、误检率高等问题,为了解决该问题,提出一种基于对比学习生成式对抗网络(ContrastGAN)的无监督检测方法。首先,建立基于ContrastGAN的色织物图像重构修复模型;其次,采用对比学习加强潜在特征空间正负例样本约束,最大化输入输出图像对应Patch之间的互信息,增强正负例特征向量区分度,使模型重构无缺陷样本图像能力得到进一步提升;然后,利用训练好的模型得到待测色织物的重构图像,并通过计算得到待测样本与对应重构图之间的残差图像;最后,对残差图像进行阈值分割和数学形态学处理,实现了缺陷区域的快速检测和准确定位。该模型能有效重构多种色织物的纹理,相比传统自编码器能够实现更高的缺陷定位精度,满足多种复杂色织物缺陷检测场景的需要。

The traditional autoencoder was used for the defect detection task of yarn-dyed fabric with complex pattern and background texture,which has problems of worse universality,higher missed detection rate and false detection rate and so on.In order to solve the problems,an unsupervised detection method based on contrastive leaning generative type of adversarial networks(ContrastGAN)was put forward.Firstly,image reconstruction and repair model of yarn-dyed fabric based on ContrastGAN was established.Secondly,contrastive learning was used to strengthen constraints of positive and negative samples in the latent feature space.Corresponding patch level mutual information between the input and output images were maximized.Discrimination of the positive and negative feature vectors was enhanced.Ability of the model to reconstruct defect-free sample images was further improved.Then,reconstructed image of yarn-dyed fabric was obtained by using the trained model.Residual images between the tested samples and the corresponding reconstructed image were obtained by calculation.Finally,rapid detection and accurate location of the defect region were achieved by the threshold segmentation and mathematical morphological processing for residual images.Texture of multiple yarn-dyed fabric could be effectively reconstructed by the model.Compared with traditional autoencoders,higher defect localization accuracy could be achieved.And needs of defect detection scenarios for various complex yarn-dyed fabrics could be met.