Image retrieval using deep saliency edge feature

Zhou Lu, Guanghai Liu, Zuo-Yong Li, Bo-Jian Zhang

Abstract: Exploiting a compact representation has become a hotspot in deep learning-based image retrieval. Fine-tuning the deep networks can provide the high discriminating representations, but it is very difficult to obtain sufficient labeled data. However, using the unsupervised methods to provide efficient representations remains challenging. Therefore, we propose a compact and robust representation, namely deep saliency edge feature (DSEF), to image retrieval. Its main highlights are: (1) Color differences, spatial layout, and edge cues within various object regions are combined into saliency edge feature maps. It can reflect a large amount of discriminative information contain in deep feature maps, thereby improving the discriminative power of deep features. (2) Edge cues are utilized to highlight the rough targets contained in deep feature maps. It can reduce the semantic disconnect exists in the different kinds of features and promote the compatibility of deep features and handcrafted features, thereby providing convenience to combine them. (3) A feature aggregation method, namely crucial cues aggregation, is proposed to aggregate crucial cues hidden inside handcrafted features and deep feature maps into a compact and high discriminating representation. Comparative experiments demonstrated that our method has provided the outstandingly retrieval performance on some benchmark datasets. The mean average precision of our method is 2.9%, 4.6%, 3.2%, 6.0% and 3.7% higher than that of most methods on the Oxford5K, Paris6K, Oxford105K, Paris106K and Holidays datasets, respectively.

External IDs:dblp:journals/eaai/LuLLZ25