Multi-Scale Anomaly Representation Learning for Robust Visual Defect Detection in Industrial Manufacturing

Abstract

Automated visual defect detection plays an increasingly important role in intelligent manufacturing by improving product quality and reducing dependence on manual inspection. Nevertheless, industrial inspection systems continue to face challenges associated with limited defective samples, subtle surface abnormalities, variations in material textures, and changing imaging conditions. This study proposes a multi-scale anomaly representation learning framework for robust visual defect detection in industrial manufacturing environments. The proposed approach combines hierarchical visual feature extraction, cross-scale feature aggregation, and normal-pattern reconstruction to identify structural and textural anomalies without requiring extensive defect annotations. A feature consistency optimization strategy is introduced to improve the separation between normal and anomalous representations, while an adaptive anomaly scoring mechanism enhances sensitivity to small and low-contrast defects. The framework is designed for evaluation using the MVTec AD dataset, containing 5,354 high-resolution images across 15 industrial object and texture categories. Comparative experiments will include Autoencoder-based anomaly detection, PatchCore, PaDiM, and conventional convolutional classification approaches. Model performance will be evaluated using image-level AUROC, pixel-level AUROC, anomaly localization accuracy, and inference latency. Additional experiments will examine detection robustness under synthetic image noise, illumination variations, and limited training data. The proposed framework aims to provide an efficient and generalizable solution for automated quality inspection in industrial production systems.