Steel Defect Detection Using a ResNet-Inspired Model and Fourier Transforms

Defect detection in steel materials is a critical task to ensure product quality and reliability in various industrial applications. Traditional methods for detecting defects in steel surfaces are often labor-intensive, time-consuming, and prone to subjectivity and errors. However, the advent of deep learning-based models and techniques has provided a promising avenue for automating this process, significantly reducing error rates. In this research, the authors propose a Steel Defect Detection system with the help of a custom designed deep neural network inspired by the ResNet architecture. Notably, the model incorporates innovative attention layers not previously integrated into similar architectures, enhancing its predictive capabilities. The authors make use of data augmentation techniques to improve the model performance to generalize, and detect defects in imperceptible data. The novel multiclass semantic segmentation model has achieved over 91% accuracy and can be used to automate the Steel defect detection process, which can remarkably decrease the costs and time associated with manually inspecting the Steel sheets.