Feature disentangling and reciprocal learning with label-guided similarity for multi-label image retrieval

Image retrieval usually faces scale-variance issues as the amount of image data is rapidly increasing, which calls for more accurate retrieval technology. Besides, existing methods usually treat pair-image similarity as a binary value which indicates whether two images share either at least one common label or none of shared labels. However, such similarity definition cannot truly describe the similarity ranking for different numbers of common labels when handling the multi-label image retrieval problem. In this paper, a Feature Disentangling and Reciprocal Learning (FDRL) method is introduced with label-guided similarity to solve the above multi-label image retrieval problem. Multi-scale features are first extracted by BNInception network and then disentangled to the corresponding high-and low-correlation features under the guidance of estimated global correlations. After that, the disentangled features are combined through a reciprocal learning approach to enhance the feature representation ability. Final hash codes are learned based on the global features derived from BNInception network and the combined features generated by reciprocal learning. The whole network is optimized by the proposed label-guided similar-ity loss function which aims to simultaneously preserve absolute similarity for hard image pairs and rel-ative similarity for soft image pairs. Experimental results on three public benchmark datasets demonstrate that the proposed method outperforms current state-of-the-art techniques. The code is online here: 'https://github.com/Yong-DAI/FDRL'.(c) 2022 Elsevier B.V. All rights reserved.

Automated summary

This paper introduces a Feature Disentangling and Reciprocal Learning (FDRL) method with label-guided similarity to solve the multi-label image retrieval problem. It enhances the feature representation ability through feature extraction, disentanglement, and reciprocal learning, and optimizes the whole network using a label-guided similarity loss function. Experimental results show that the proposed method outperforms current state-of-the-art techniques.