Saliency Model Based on Discriminative Feature and Bi-directional Message Interaction for Fabric Defect Detection

• Published in: Digital TV and Wireless Multimedia Communication pp. 181 - 193
• Main Authors: Liu, Zhoufeng, Wang, Menghan, Li, Chunlei, Guo, Zhenduo, Wang, Jinjin
• Format: Book Chapter
• Language: English
• Published: Singapore Springer Singapore
• Series: Communications in Computer and Information Science
• Subjects: Discriminative feature; Fabric defect detection; FCN; Saliency model
• ISBN: 9811611939; 9789811611933
• ISSN: 1865-0929; 1865-0937
• DOI: 10.1007/978-981-16-1194-0_16

Due to the complexity of the texture background and the diversity of the defect types of the fabric image, the traditional method of fabric defect detection shows poor detection performance. Recent advances on salient object detection benefit from Fully Convolutional Neural Network (FCN) and achieve the good performance. Meanwhile, the deficiencies in a fabric image are salient compared to the texture background, so the saliency model is very feasible for fabric defect detection. In this paper, we propose a novel saliency model based on discriminative feature and bi-directional message interaction for fabric defect detection. Firstly, we design a multi-scale attention-guided feature extraction module in which the multi-scale context-aware feature extraction block and channel attention block are respectively used to capture multi-scale contextual information and assign greater weight to more discriminative features corresponding to the right fabric defect scale. Then a bi-directional message interaction module is designed to promote the effectiveness of feature with specific resolution by interacting message along both directions, which further improves the availability of the feature extraction. After the bi-directional message interaction module, we use a cross-level contrast feature extraction module, which elevates features with locally strong contrast along each resolution axis, to predict saliency maps. Finally, the predicted saliency maps are efficiently merged to produce the final prediction result. We conduct extensive experiments to evaluate our net and experiment results demonstrate that the proposed method outperforms the state-of-the-art approaches.