Next-Generation Defect Intelligence: Leveraging Xception Feature Embeddings with Deep Convolutional Inference for Surface Anomaly Detection

Abstract

Metal surface quality inspection from line scan images has become a critical requirement in modern manufacturing, with studies indicating that nearly 20 to 25 percent of total production waste arises from undetected surface defects, and automated inspection systems have shown up to a 40 percent improvement in detection consistency over manual methods. The increasing demand for high precision in sectors such as automotive, aerospace, and steel processing highlights the need for fast, reliable, and scalable inspection systems capable of identifying microscopic defects under continuous production flow. Traditional manual inspection often suffers from operator fatigue, inconsistent judgments, and limited ability to detect subtle or high-frequency defects in real time, making it unsuitable for high-volume industrial environments. To overcome these challenges, this work proposes a deep learning-based inspection framework trained on the Northeastern University (NEU) Metal Surface Defects Data, integrating deep depth wise separable convolutions to achieve lightweight yet high resolution feature extraction. While conventional algorithms such as random forest, naive bayes, decision tree, logistic regression, and classical convolutional neural networks (CNN) struggle with complex texture variations, the proposed method leverages the Xception with CNN architecture to enhance feature separation, reduce computation, and improve defect classification robustness. The system outputs six major defect categories observed in metal manufacturing: crazing, inclusion, patches, pitted, rolled, and scratches, offering an efficient, scalable, and industry ready solution for automated quality control.