Small-scale block defect detection of fabric surface based on SCG-NET.

In contrast to the common small target detection problems, it is more difficult to locate and identify the small surface defects of fabric due to its own texture and complex background interference. Therefore, this paper proposes an effective detector for small-scale block defects on fabric surface by taking advantage of the backbone which integrates the Coordinate Attention module to enhance the acquisition of small-scale block defect location information. The FPN + PAN multi-scale detection structure is adopted to effectively integrate the feature information between different levels and deal with the multi-scale problem of defects. In the Neck section, a small target detection layer is set to expand the receptive field to prevent the loss of small-scale defect feature information. Moreover, we propose to use the GhostBottleneck module instead of the ordinary downsampling process to eliminate redundant convolutional calculations to improve the detection speed. The experimental results show that the optimal detection results of 0.56 and 0.842 are achieved in the detection recall and accuracy of the public fabric dataset; compared with other detectors, the result of small-scale defect detection rate is reduced by at least 2.7%, and the detection process meets the real-time requirement of automatic defect detection, which verifies the effectiveness of our method. Code and data are available at: https://github.com/VIMLab-hfut/SCG-NET. [ABSTRACT FROM AUTHOR]