基于多尺度卷积自编码器的色织面料缺陷检测

Defect detection of yarn-dyed fabric based on multi-scale convolutional auto-encoder

针对传统自编码器对复杂花型色织衬衫面料缺陷检测效果不佳的问题,提出一种基于多尺度卷积自编码器的色织衬衫面料缺陷自动检测与定位方法。对无缺陷样本集加入椒盐噪声,建立多尺度卷积自编码器重构模型,该模型将特征映射通过跳接的方式连接起来进行多尺度特征融合;然后,将添加噪声后的样本放入模型训练,使模型具有对噪声干扰进行重构性修复的能力;计算待测色织衬衫面料图像和其重构图像的残差;将所得残差图像进行数学形态学处理,实现色织衬衫面料缺陷区域的检测和定位。实验结果表明,所提出的算法在不依赖样本标注的情况下,可以实现复杂花型色织衬衫面料缺陷区域的检测和定位。

In order to solve the problem that the traditional auto-encoder was not effective in detecting the defects of complex pattern yarn-dyed shirt fabrics,a method of automatic detection and location of the defects of yarn-dyed shirt fabrics based on multi-scale convolutional auto-encoder was proposed.Salt and pepper noise was added to the defect-free samples,a multi-scale convolutional auto-encoder reconstruction model was established,and the feature maps of the two parts of the network were fused with multi-scale features through transverse connection.The samples with the added noise were put into the model for training,so that the model has the ability to reconstruct the noise interference.The residual image between the input image and its reconstructed image was calculated.The residual image was processed by mathematical morphology,and the defect area was detected and located.The experimental results show that the proposed algorithm can detect and locate the defect area of complex pattern yarn-dyed shirt fabric without relying on the labeled sample.