SCNet: a novel network involving SIFT for cross-scenario fabric image retrieval

Cross-scenario fabric image retrieval faces challenges of texture deformation, illumination variations, partial occlusions, and intra-class appearance differences. To improve the feature representation ability, this study proposes a novel network SCNet (SIFT-guided Cross-scenario Network) for cross-scenario fabric image retrieval. SCNet involves hand-crafted features to avoid scheme design and parameter optimization for end-to-end feature representation. The hand-crafted features that are robust to cross-scenario changes are selected as the prior knowledge to guide and accelerate the model training. The multi-level features are automatically captured by the designed multi-granularity supervised loss functions and model training strategies. Experimental results prove that the proposed method can combine hand-crafted features and semantic features, being effective and superior for cross-scenario fabric image retrieval. The proposed scheme can provide the references for the user in the factory and e-commerce platform, saving lots of labor and material resources.