Abstract
Because of the complex and diverse fabric image texture and defects, the traditional fabric defect detection algorithm has poor detection results and low efficiency. Visual saliency model can outstand the defect region from the complex background. However, the previous saliency detection models typically utilize hand-crafted image features to generate the saliency map, and it can only be used for some kinds of fabric type. In this paper, a deep saliency model generated by fully convolutional network with attention mechanism is proposed for fabric defect detection. First, the proposed model extracts multi-level and multi-scale features using Fully Convolutional Networks (FCN), this will improve the characterization ability for fabric texture. Then, the attention mechanism module is incorporated into the backbone network, thus the different feature map is assigned different weight, this further improves the effectiveness of the feature extraction. Finally, multi-level saliency maps are generated after deconvolution, and then fused by a series of short connection structures to better detect the salient region. Experiment results demonstrate that the proposed approach can accurately locate the defect region comparing with the state-of-art methods. Meantime, defect detection ability of the network model can be improved without significantly increasing the amount of calculation and parameters.
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have