Surface defect inspection is vital for ensuring high-quality manufacturing. Surface defects recognition, based on Convolutional Neural Networks (CNNs), have recently demonstrated remarkable success in spotting anomalies. However, a significant limitation is that they rely on an ample quantity of training defect samples, which cannot be easily collected in real-world scenarios. Consequently, we proposed joint Frequency Fusion and Multi-Channels Attention Neural Network ( \(\text {F}^2\) MCANet) to enhance the utilization efficiency of features, thereby improving surface defect recognition performance even with limited labeled samples. Firstly, the Multi-Channels Attention (MCA) module utilizes both max pooling and average pooling to endow the network with the capability to process global and local information across different channels. Additionally, we introduced a Frequency Fusion (FF) module designed to leverage the distinct frequency information available at different levels. This module facilitates the network in learning both low (e.g., semantic) and high (e.g., texture) frequency features within the constraint of limited supervision data, enabling more comprehensive recognition of defect images. Our method outperformed state-of-the-art CNNs while its reliability and effectiveness were verified by extensive ablation studies.
