中国科学院自动化研究所机构知识库

Constructing textile defect detection systems is significant for quality control in industrial production, but it is costly and laborious to label sufficient detailed samples. This paper proposes a model called 'spatial adversarial convolutional neural network' which tries to solve the problem above by only using the image-level label. It consists of two parts: a feature extractor and feature competition. Firstly, a string of convolutional blocks is used as a feature extractor. After feature extraction, a maximum greedy feature competition is taken amongfeatures in thefeature layer. The feature competition mechanism can lead the network to converge to the defect location. To evaluate this mechanism, experiments were carried on two datasets. As the training time increases, the model can spontaneously focus on the actual defective location, and is robust towards an unbalanced sample. The classification accuracy of the two datasets can reach more than 98%, and is comparable with the method of labelling samples in detail. Detection results show that defect location from the model is more compact and accurate than in the Grad-CAM method. Experiments show that our model has potential usage in defect detection in an industrial environment.