基于记忆去噪卷积自编码器的色织物缺陷检测

Defect detection of yarn-dyed fabric based on memory denoising convolutional auto-encoder

针对传统自编码器泛化能力弱导致色织物缺陷检测性能不佳的问题,提出一种记忆去噪卷积自编码器重构模型和残差分析的无监督色织物缺陷检测与定位方法。首先,训练阶段仅利用无缺陷样本叠加椒盐噪声构建训练集。接着,建立记忆去噪卷积自编码器重构模型。然后,将训练集输入模型进行训练,使模型具有重构修复缺陷区域的能力。最后,在检测阶段计算待测色织物图像和其对应的重构图像之间的残差,并对残差图像进行阈值分割和闭运算操作,实现色织物缺陷区域的检测和定位。实验结果表明,提出的方法能有效重构色织物纹理,快速准确地检测和定位多种色织物的缺陷区域。该方法无需缺陷样本和缺陷样本标记,仅通过记忆无缺陷样本特征来增强模型重构修复缺陷区域的能力,从而提高缺陷检测性能。

Aiming at the problem of the weak generalization ability of traditional auto-encoder leading to poor performance of yarn-dyed fabric defect detection,a memory denoising convolutional auto-encoder reconstruction model and residual analysis unsupervised yarn-dyed fabric defect detection and location method was proposed.First,only defectfree samples were added with salt and pepper noise to construct a training set in the training stage.Second,a reconstructed model of memory denoising convolutional autoencoder was established.Then,the training set was input into the model for training,so that the model has the ability to reconstruct and repaired the defect area.Finally,the residual map between the tested fabric image and correspondingly reconstructed image was calculated in the detection stage,and the threshold segmentation and closing operation were carried out on the residual map to realize the detection and location of the yarn-dyed fabric defect area.Experimental results show that the proposed method can effectively reconstruct the texture of yarn-dyed fabric,quickly and accurately detect and locate the defect areas of yarn-dyed fabrics.This medthod does not require defective samples and defective sample labels,but improves the defect detection performance by memorizing the characteristics of defect-free samples to enhance the ability of model reconstruction to repair the defect area.