Textile Defect Detection Algorithm Based on Unsupervised Learning

In order to solve the problem of lack of sample data and high cost of dataset due to the large number of abnormal samples and high-precision marking data required by current deep learning algorithms in textile defect detection, a pixel-level real-time defect detection scheme based on autoencoder and morphology was proposed in this paper. Algorithm is innovation in that can carry on the network training by unsupervised learning, as opposed to supervised learning needs a large number of high-precision marking abnormal samples, the algorithm relies on the dataset is only normal sample data, and no need to tag samples. In addition to reducing the production cost of large dataset, textile defects of various sizes can be detected in real-time at the pixel level. The algorithm steps are described as follows: First, the normal textile image is input into the network for encoding and decoding, and the underlying feature information of textile image is learned and reconstructed into a new image. Secondly, the encoding and decoding stages were combined horizontally to obtain better fitting effect. By subtracting the input image from the reconstructed image, the difference matrix of the input image and the reconstructed image was obtained to obtain the range of the defect area. Finally, Dilate, Median Filtering and Edge Detection are used to amplify and denoise the features of the defect region to obtain the final accurate defect region. The experimental results show that the scheme can effectively detect textile defects in real-time at pixel-level only when normal samples are used as dataset. Compared with supervised learning based algorithms such as RCNN and YOLO, this scheme only needs normal samples as dataset to carry out network training, which greatly reduces the cost of making dataset. Besides, Accuracy and F1-score can both reach over 0.95 in 4 different textile datasets, and its FPS is 36.2. Meet the requirements of real-time detection. The code and models will be made publicly available at https://github.com/hanknewbird/anomaly-detection.