Fabric image retrieval in complex scenes based on hand-crafted features and feature fusion

With the rapid expansion of image databases and the increasing diversity of data sources, the same fabric may be captured under varying folding and curling conditions, as well as different background settings. To enhance the efficiency of fabric image retrieval in complex scenarios, we propose a method that integrates global texture features with multiscale color histogram features, enabling customers and manufacturers to quickly and accurately retrieve identical or similar products from the database. First, median filtering is applied to reduce noise and suppress irrelevant background interference. Next, local features are extracted using the scale-invariant feature transform (SIFT) algorithm, and these features are further aggregated into global texture descriptors via the vector of locally aggregated descriptors (VLAD) method. Meanwhile, a multiscale color distribution extraction method is designed based on the color histogram. The histograms at different scales are normalized and concatenated to form the final color feature representation. Third, the similarities of color features and textures are measured using corresponding distance functions. To evaluate the proposed scheme, a new fabric image retrieval dataset containing over 28,000 images is constructed as a benchmark. Experimental results demonstrate that the framework effectively improves both the precision and efficiency of image retrieval in complex scenarios. This method provides a practical and effective solution, offering valuable references for industry and customers while enhancing retrieval efficiency.