An efficient CNN-Mamba hybrid network for fabric defect segmentation

Fabric defect detection is crucial for quality control and economic efficiency in the textile industry. Traditional methods and CNN-based approaches struggle with complex textures and high-frequency defect features. Although Transformers excel in capturing the global context of images, their high computational demands hinder their deployment in industrial settings. The recently emerged Mamba model has shown great potential in computer vision due to its ability to model global context with linear complexity. However, we have observed a limitation in the pure Mamba model, which exhibits a bias towards low-frequency features, potentially overlooking critical high-frequency defect details. In response, We introduce a novel CNN-Mamba hybrid model with a Wavelet Feature Decomposition (WFD) module. The WFD uses Haar Wavelet Transform to decouple features, enabling Mamba to process low-frequency information and CNNs to enhance high-frequency details. Our method achieves superior performance with fewer parameters, as demonstrated through experiments on three fabric datasets. • Proposes a novel CNN-Mamba hybrid architecture for fabric defect segmentation with enhanced feature extraction capabilities. • Introduces a Wavelet Feature Decomposition module module to address low-frequency bias in Mamba, achieving effective frequency decoupling for improved performance. • Demonstrates superior defect identification and localization across diverse fabric textures through extensive experiments on multiple datasets.