Abstract:

Objective Fabric image retrieval is useful to search for similar fabric images already existing in a factory to obtain corresponding process parameters to guide production, and save manpower and material resources in repeated trial weaving process. The existing fabric image retrieval methods have not combined the characteristics of plaid fabric images, and the retrieval performance needs further improvement.

Method To improve the accuracy of plaid fabric image retrieval, a convolutional neural network (CNN) model that combines attention mechanism and hash coding was proposed based on the characteristics of plaid fabrics. By introducing attention mechanism into the CNN branch to focus on key information of fabric images, global and local depth features were extracted and fused by the feature fusion module. A hash encoding layer was built to compress the fused features to balance the retrieval precision and retrieval efficiency.

Results The proposed method was tested by establishing a new plaid fabric image retrieval dataset. Results showed that the average precision (P₅) and recall(R₅) of four categories were up to 77.3% and 55.2%, respectively, indicating the system's capability of retrieving the relevant images effectively. The average mean average precision (mAP) of 0.759 demonstrated the robustness of the system across all queries. In addition, there was a positive correlation between P₅, R₅, and mAP. The retrieval precisions of color block and academic were up to 83.1% and 86.5%, respectively. The retrieval precisions of the window and Welsh were relatively low, being 71.8% and 67.4%, respectively. The reason was that some window and Welsh plaids had a cycle period exceeding the image acquisition size of this study, which resulted in differences between images of the same fabric sample collected in different areas. The average P₅ reaches 77.3%, suggesting that 3.865 out of the ranked top-5 images are correct. The requirement for enterprises to conduct fabric image retrieval is to quickly find similar images, and the proposed method was proven to meet the requirements of fabric manufacturing enterprises. Different low-order feature extraction methods and deep feature extraction methods were compared with the proposed method. Various types of plaid fabrics were with significant intra-class and inter-class differences. These low-level features heavily relied on manual design and the parameters were not universally applicable for different types of plaid fabrics. Although deep features could achieve good performance, the retrieval performance was still lower than the proposed method due to the lack of integration with the characteristics of plaid fabric images. The comparative experiments proved the adaptability and superiority of the proposed method for plaid image retrieval.

Conclusion A plaid fabric image retrieval method based on deep feature fusion was proposed. The attention mechanisms and hash encoding layer were involved into the CNN models to realize feature extraction, feature fusion and feature compression. Plaid fabric image retrieval was achieved by similarity measurement using the hamming distance. The results showed that the average P₅, R₅, and mAP could reach 77.3%, 55.2%, and 0.759, respectively, demonstrating the feasibility and effectiveness of the method. By comparing existing fabric image retrieval methods, the adaptability and superiority of the proposed method for plaid image retrieval are verified. The proposed method can help enterprises quickly search for existing product process parameters to guide production and improve design, production, and operational efficiency.

Key words: deep feature, convolutional neural network, attention mechanism, hash encoding, fabric image retrieval, plaid fabric image, production efficiency