ACR Net: Adaptive Contrast Reconstruction U-Net for Hot-Rolled Steel Defect Detection

In the field of defect detection in hot-rolled strip steel, deep learning methods have emerged as a hot research topic and have been widely applied in practical production, renowned for their outstanding detection accuracy. In practical production, many defects on the surface of hot-rolled strip steel exhibit low contrast with the background due to factors such as insufficient illumination, placement of image acquisition devices, and smoke and dust. This low contrast situation makes it difficult for deep neural networks to extract features from grayscale images, thereby reducing the performance of defect detection. To address this issue, this paper proposes a novel network called ACR-Net (Adaptive Contrast Reconstruction Net) aimed at enhancing defect detection performance under low contrast conditions. ACR-Net consists of an ACR U-Net sub-network, a dual-branch feature extraction network, and dense feature fusion. ACR U-Net achieves contrast reconstruction and high-contrast information extraction by incorporating a global dynamic adaptive contrast reconstruction layer (ACR layer), the dual-branch feature extraction network enhances the network's robustness, and the dense feature fusion (DFF) progressively fuses hierarchical features. Experimental results demonstrate that on the NEU-Seg datasets and SUT-seg datasets, the ACR-Net method achieves an average intersection over union (mIoU) of 82.81% and 65.30%, respectively, outperforming existing methods.