Surface Defect Detection on Steel Plates Based on Improved Unet Network

In the steel manufacturing process, various defects may arise in steel plates due to inadequate manufacturing processes. To address issues such as low accuracy and missed detections during the inspection process, this paper proposes an enhanced Unet model for surface defect detection. Firstly, the model incorporates the Global Context (GC) module, replacing traditional skip connections, thereby enhancing the model's perception of features at different scales and levels to improve segmentation performance. Secondly, to tackle the challenges of sample imbalance and uneven defect areas, a composite loss function is designed for training, comprising Focal loss, Tversky loss, and BCEWithLogitsLoss. This addresses sample imbalance and the difficulty in segmenting small regions. The results demonstrate that our model outperforms traditional Segnet, Fcn, and Unet models in terms of accuracy and segmentation metrics, including IOU and Dice coefficients.