Perception of flaws in Steel using ResUNet Algorithm

Manual inspection techniques are still utilized in conventional industrial manufacturing to find product surface flaws because of the limits of science and technology. The current, well-liked deep learning-based approach has a poor level of precision and has a lot of room for improvement. Due to the constraints of manual labour and outdated technology, this process is sluggish and ineffective. The purpose of this thesis is to investigate if automating this utilizing cutting-edge computer technology and picture flaw categorization using one of the recently developed deep learning techniques is feasible. We suggest ResUNet, a mix of U-net and ResNet50 architecture, as a step towards creating a fully automated model for pixel-wise segmentation that can detect steel defects in images. Our experimental findings indicate that the recommended design yields successful segmentation outcomes on datasets that are made available to the general public. The paper comes to the conclusion that it is feasible to attain a dice coefficient of more than 76% based on the outcomes of controlled testing.