Metal Surface Defect Classification Using Custom Convolutional Neural Network with Self Attention Mechanism

Abstract

The metal surface of machinery plays an indispensable role in numerous industries, in today's bustling environment. A defective metal surface lowers the performance and quality of the product. The surface examination is often done manually or using crude automated techniques, but the issue cannot be entirely resolved due to external variables. This study introduces an innovative approach to identify metal surfaces by pre-processing, leveraging deep learning methodologies, specifically integrating Convolutional Neural Networks (CNN) with self-attention mechanisms. It aims to enhance the ability of the model to recognize minute details that are essential for precise detection. The self-attention modules enable the network to capture complex patterns and spatial relationships by choosing focusing on specific areas within metal surface dataset, enhancing sensitivity to difficult conditions.