SFCNet: A Siamese feature contrast network for surface defect segmentation on printed circuit boards

In recent years, deep-learning-based methods have been widely applied in the field of industrial surface defect detection, including the precise segmentation of defects on printed circuit boards (PCBs). However, the high time and financial costs associated with obtaining trainable annotated datasets in practical production severely limit the performance of deep-learning-based segmentation networks. To address this issue, we propose a Siamese feature contrast network (SFCNet) that utilizes easily obtainable template images to enhance the performance of segmentation networks. The proposed SFCNet comprises two shared-weight encoders, a decoder, a feature contrast module (FCM), and an edge feature module (EFM). The shared-weight encoders are designed to extract multilevel high-level features from both template and input images. The FCM and EFM extract different pieces of information from the features and local edge features of the two images, respectively, and the fused features from these two modules are fed into the decoder to compute the defect prediction map. Unlike existing Siamese or contrastive learning-based methods that focus primarily on global feature similarity, SFCNet introduces specialized modules to explicitly capture both multiscale contrastive differences and fine-grained edge cues, improving defect localization and segmentation precision. Notably, we observe that using higher-level features in the FCM could degrade network performance, which we attribute to the abstract information in higher-level features not aiding the proposed FCM. In addition, we introduce a composite loss function for network training, which combines Lovász-Softmax Loss and Ohem Cross-Entropy Loss to calculate the loss between predictions and labels, employing weighted values to achieve the best combination. Extensive experimental results on real PCB surface defect datasets demonstrate that SFCNet achieves significant performance improvements over various baselines. For instance, when integrated with SFCNet, DANet++, and ANNet++ achieve intersection over union (IoU) score improvements of 8.61% and 6.05%. ANNet++ reached an IoU score of 64.12%, which is the highest among all evaluated advanced models.