Abstract:

Objective In the textile industry， fabric retrieval invokes directly relevant technical parameters to guide the production process through inquiring existing similar products. The procedures of sample analysis and repeated trial weaving are reduced significantly， achieving digitized and intelligent management. The current fabric retrieval methods leave out the low-level visual information of various plaid fabrics and high-level semantic information including lattice and style， which fail to meet the accuracy requirement of retrieval in the segmentation of plaid fabrics.

Method The low-level features of plaid fabric images were characterized by the designed local texture features， key-point texture features， local color features， and spatial color features. Meanwhile， attention mechanisms were introduced into existing CNN network models to extract global and local depth features， and feature fusion and hash encoding were performed to realize efficient search. The similarities of different features were measured based on the distance function， and the weight allocation was used to combine low-level and high-level features.

Results A new plaid fabric image retrieval dataset containing 44 000 images was built as the benchmark to evaluate the proposed method. Results showed that the average precision at top 5 （P₅）， recall at top 5 （R₅） and mean average precision （I_mAP） of the four categories（including solid-color grids， window grids， academic grids， and Welsh grids） reached 79.6%， 59.5%， and 0.780， respectively， verifying the feasibility and effectiveness of the proposed method. The 79.6% precision P₅ of the top 5 images means that about 3.95 of the top 5 images were highly correlated with the required contents. In terms of visual similarity， the proposed method effectively retrieved images with similar textures and colors， which were highly similar to the query image in both global appearance and local details. Compared with single low-level and high-level feature-based retrieval performance， the retrieval metrics P₅of low-level and high-level feature combination method was improved by 5.2% and 2.1%， R₅ was improved by 6.9% and 2.4%， and I_mAP was improved by 0.062 and 0.022， respectively. From the perspective of improvement effect， the combination of low-level and high-level features was able to enhance effectively the retrieval performance of plaid fabric images， leveraging the advantages of different features to form complementary advantages. Compared with the existing image retrieval methods， the results of the new method suggested the adaptability and superiority for plaid image retrieval.

Conclusion This paper proposes a novel plaid fabric image retrieval method based on low-level and high-level feature combination. The retrieval of plaid fabric images has been achieved by integrating different features through weight allocation to form complementary advantages. The results showed that the average P₅， R₅ and I_mAP of four categories can reach 79.6%， 59.5%， and 0.780， respectively， demonstrating the feasibility and effectiveness of the method. The comparative experiments prove the adaptability and superiority of the proposed method for plaid image retrieval. The proposed method can provide reference for textile enterprises to search for the required fabric images and improve their design， production， and operational efficiency.

Key words: low-level feature, high-level feature, feature combination, plaid fabric, image retrieval