Adversarial-based refinement dual-branch network for semi-supervised salient object detection of strip steel surface defects

The detection of surface defects on strip steel poses significant challenges due to factors such as background noise and interference. Moreover, the lack of labeled defect data in practical scenarios makes it difficult to effectively differentiate defects using computer vision and machine-learning methods. To tackle these issues, we propose a novel semi-supervised saliency detection method called adversarial-based refinement dual-branch network (ARDNet). We introduce an adversarial learning mechanism to ARDNet, which leverages unlabeled data through a generator network and a discriminator network. The generator network's encoder employs multiple convolutional branches to extract multi-level features, while the multi-scale integration refinement module (MIRM) integrates semantic features from these branches and refines edge details. Subsequently, the decoder fuses deep features into the saliency map and utilizes a dual-branch structure to minimize interference between labeled and unlabeled inputs. Lastly, the discriminator network provides additional supervision to the generator network by distinguishing between the predicted probability maps generated by the generator and the ground truth segmentation distributions. Experiments on the publicly available dataset show that our approach surpasses other competitive methods in terms of filtering background noise, adapting to multi-scale defect sizes, and preserving defect details. By effectively utilizing unlabeled images, our method enhances segmentation accuracy, accurately locates defect positions, and successfully segments the defects.