Fig. 2
From: Siamese network with change awareness for surface defect segmentation in complex backgrounds
The pipeline of our Transformer-based change-aware defect detection network, CADNet, is designed to accept an image under inspection (NG) and a defect-free reference image (OK) for deep change modeling. The contrastive feature encoder, comprising a Siamese four-stage Transformer, generates a deep feature distance map (DistMap). The change-aware decoder leverages the DistMap to facilitate accurate defect localization. The network is trained using a cross-entropy loss and a multi-class balanced contrastive loss. Note that the ground truth used is a multi-class segmentation map.
