Abstract: The fabric defect detection with periodic pattern generally needs to calculate period, which is not suitable for pure texture fabric. In this paper, the detection method for two kinds of fabric is proposed. Firstly, the size of detect block is set, the image block is extracted randomly from non-defect images, and then the auto-encoder is trained. Secondly, according to the size, the test images are divided into several blocks, and reconstructed with the auto-encoder, and then the mean square errors are computed between the reconstructed result and original data. Lastly, outliers of computing results of test images are detected. With rather large mean square errors value, the blocks are defect blocks. The experiments show that the proposed method is suitable for two kinds of fabric, the process is more practical and the detection results are better.

Key words:  partitioning, auto-encoder, mean square error, defect detection