MCF-Net: a multi-scale context fusion network for real-time fabric defect detection

Fabric defect detection plays a critical role in textile manufacturing to ensure product quality. However, challenges such as significant variations in defect sizes result in poor detection performance of existing models. To address these challenges, this paper proposes a Multi-scale Context Fusion Network (MCF-Net). First, The Multi-scale Context Diffusion Fusion Pyramid Network (MCD-FPN) was proposed, which reconfigures the traditional feature fusion path and designed a key component, the Multi-scale Context Aggregation Module (MCAM). MCAM takes multi-scale features extracted from the backbone network as input and establishes long-range dependencies between features and their surrounding background through a series of deep convolutional operations. The output of MCAM is then used to enhance both shallow and deep features, ensuring that each detection scale contains comprehensive multi-scale context information. Furthermore, to improve the network’s feature extraction capability, the Latent Feature Transformer (LFT) was proposed, which maps input features into a high-dimensional space and extracts depth features through high-dimensional information compression. Second, to improve the network’s multi-scale perception, the Local Cross Attention Mechanism (LCA) was proposed, which models the spatial information of the features to improve the network’s understanding of global context information. Experimental results on the self-built FD6052 dataset, as well as the publicly available TianChi and DAGM2007 datasets, demonstrate the effectiveness of MCF-Net, which outperforms existing methods in fabric defect detection. In addition, MCF-Net achieves an inference speed of 85.6 FPS, enabling real-time detection at industrial scale.