Enhanced Steel Defect Detection using U-Net Model with Attention Mechanism

In this research, detection of steel defects is a significant use of computer vision that can enhance industrial quality control. New prospects for automated defect segmentation are presented by recent developments in deep learning. This paper investigated the use of attention processes in a U-Net convolutional neural network architecture for improved steel flaw identification. Images of steel and matching defect masks from the Severstal Steel Defect Detection dataset were used to train the U-Net network. Preprocessing actions such as image rotation and scaling were performed in addition to data augmentation. When tweaked at various probability thresholds, the model produced defect mask predictions on a test set with good accuracy. When compared to ground truth masks, qualitative results showed that different defect types could be segmented well. Forecasts on fresh test pictures without any masks provided also shown a believable ability to identify flaws. According to this research, deep learning using attention blocks and U-Nets has the potential to improve automation in tasks related to semantic segmentation, such as steel inspection. To assess the model's efficacy on other defect types and steel materials, more work is required. The code and trained model serve as a basis for creating systems that are ready for production.