Frequency-Aware Convolutional Attention for Industrial Defect Image Classification

Abstract

Industrial defect image classification requires models that can distinguish subtle texture changes, small structural abnormalities, and low-contrast surface defects. Conventional convolutional neural networks can extract hierarchical features, but they may lose fine-grained information during pooling and downsampling. This article proposes a frequency-aware convolutional attention model for industrial defect classification. The framework introduces multi-scale wavelet transform convolution to capture high-frequency defect patterns and low-frequency shape information simultaneously. A convolutional block attention module is then used to refine feature maps by learning channel-wise and spatial importance. The proposed model is suitable for surface defect detection in metal, textile, electronic component, and manufacturing inspection scenarios. The article argues that integrating frequency-domain analysis with deep learning improves recognition reliability under noise, uneven illumination, and background clutter. Experimental design focuses on classification accuracy, robustness, interpretability, and computational efficiency. The study provides a practical direction for intelligent visual inspection systems in industrial quality control.