A Multi-feature Integration Descriptor for Instance Image Retrieval

Deep features can provide remarkable performance in image retrieval, but their domain adaptability and robustness are still going along. To our knowledge, some classical visual features are robust to changes in orientation, scale, and transformation. It is still challenging to combine their advantages. To address this issue, we propose a new multi-feature integration descriptor for instance image retrieval. The main highlights are threefold: (1) we propose a new method to obtain the salient region map that exploits the luminance information of L*a*b* color space and Tamura textural features together with deep features. (2) Based on the salient region map, a spatial fusion scheme is proposed to integrate multiple features, which can highlight the object region and suppress irrelevant regions. (3) A new channel balancing strategy is introduced to enhance channels with necessary or discriminative patterns and weaken channels with redundant patterns in deep feature maps. Detailed experimental results show that our proposed method outperforms several recent state-of-the-art methods on five public datasets.