Lin, D., Li, Y., Prasad, S., Nwe, T. L., Dong, S., & Oo, Z. M. (2021). CAM-guided Multi-Path Decoding U-Net with Triplet Feature Regularization for Defect Detection and Segmentation. Knowledge-Based Systems, 228, 107272. doi:10.1016/j.knosys.2021.107272
Abstract:
Automated defect detection and segmentation from high-resolution industrial images is an essential and challenging task. In this paper, we design a novel CNN network called Class Activation Map Guided U-Net (CAM-UNet) to address this task. The proposed network can be trained under the real-world industrial condition that sufficient normal (defect-free) images and a small number of annotated anomalous images are available. Technically, we first modify and pretrain the encoder of a VGG-16 backboned U-Net to classify normal and anomalous images. After pretraining, the class activation maps (CAMs) can be generated as the guidance to localize the defective regions within anomalous images. Secondly, we propose a novel Triplet Feature Regularization (TFR) module to facilitate the encoder network to simultaneously generate consistent representations of normal regions and discriminative representations between normal and defective regions. Finally, we propose a multi-path decoding (MPD) module consisting of multiple decoding subnetworks. The subnetworks are trained by minimizing three different segmentation losses and their outputs are aggregated to generate the predicted defective masks. Extensive experiments are conducted on the publicly available industrial datasets MVTec AD and MTSD to demonstrate the superiority of the proposed method over multiple competing methods in both industrial defect detection and segmentation tasks.
License type:
Attribution-NonCommercial-NoDerivatives 4.0 International (CC BY-NC-ND 4.0)
Funding Info:
There was no specific funding for the research done