基于深度学习的生成式网络目前已被广泛运用于无监督缺陷检测任务中,它的训练过程仅需学习正常样本,就能将异常样本重建为正常样本,重建图像与异常样本之间的误差便可被用作识别异常的标准.但是在纺织品缺陷检测任务中,由于深度神经网络具有较强的泛化性,异常区域一般会被保留在重建图像中,导致异常的漏检.为此,提出了一种基于筛选增强自编码器的轻量化生成模型.首先,通过生成模型重建待测图像;其次,在确保重建图像不会保留缺陷区域的前提下,计算它与待测图像之间的残差图;最后,对残差图进行形态学处理后得到缺陷分割的结果.实验结果表明,所提算法能有效重建纺织品纹理,从而可准确分割出缺陷区域,分割准确率、召回率和F1分数分别可达到92.32%、89.11%和90.69%,并且所提算法还能在不同纹理的纺织品上快速迁移,提升了该类产品的检测效率.
Generative networks based on deep learning have been widely used in many unsupervised defect detection tasks.The training process only requires the normal sample,and the network can reconstruct an abnormal sample into a normal sample,and the error between the reconstructed image and the abnormal sample can be used as a criterion for identifying abnormalities.However,in the textile defect detection task,the abnormal region is generally retained in the reconstructed image due to the strong generalization of the deep neural network,resulting in the missed detection of abnormalities.To solve the problem,a lightweight generation model based on selecting enhanced autoencoder was proposed.Firstly,the test image was reconstructed by a generation model,then the residual map between it and the test image was calculated on the premise that the reconstructed image will not retain the defect regopm,and finally the result of defect segmentation was obtained after morphological processing of the residual map.The experimental results show that the proposed algorithm can effectively reconstruct the textile image so that the defect region can be accurately segmented.The precision,recall,and F1 score of the segmentation results can reach 92.32% ,89.11% ,and 90.69% .And the proposed algorithm is also able to quickly migrate textiles with different textures,which improves the detection efficiency of this type of product.