A study on Surface Flaw Detection using Deep Learning Approach

Surface Flaws are always characterised by human eyes. In this article we tackle the problem of surface flaws in metal structures. When surface flaws in metal structures are identified in a timely manner, it is possible to efficiently regulate the quality, wear, and safety of these structures. Numerous neural network topologies that have shown great efficiency in various domains, such as image classification, segmentation, and recognition, have been presented by researchers over the last ten years. We will try to implement a sophisticated model for use in production and maintenance departments, aiding in defect detection, classification, and localization on steel surfaces. The main aims of this project are to investigate Transfer Learning approaches and develop a Layered Deep Learning Pipeline for Classification and Segmentation using Convolutional Neural Networks (CNNs), Residual Networks (ResNet), and U-Net Architecture. The images in the collection each display different kinds of flaws and are labelled to indicate their location on the steel surface. The techniques used include image augmentation, data pre-processing, and building a reliable data generator for effective model training. A ResNet-based classifier will be refined for defect classification using the provided defect and non-defect photos. Simultaneously, a new ResUNet architecture will be created for defect segmentation, leveraging the advantages of residual connections for accurate localization. The goals of this research project are to achieve precise fault localization and high classification accuracy. We believe that the suggested methodology would significantly improve industrial fault detection procedures, resulting in higher operational effectiveness and lower costs. Through extensive experimental analysis conducted with a large dataset containing both defective and normal images, the module demonstrates better accuracy for both classification and segmentation.