Image retrieval using compact deep semantic correlation descriptors

Significant progress has been made in instance image retrieval based on deep feature aggregation. However, existing approaches are limited by two issues: 1) The inability of deep features to localize target objects generates inaccurate feature descriptions and 2) using short vector feature representations provides unsatisfactory retrieval performance. To address these issues, we propose the compact deep semantic correlation descriptor (DSCD) approach, which has three main highlights: (1) Unlike manual labeling approaches, we propose a channel semantic correlation learning method to localize target objects. This method can learn semantic correlation information between channels from an auxiliary dataset and use this information as a priori knowledge to locate target objects well. (2) We propose a hierarchical attention mechanism to integrate multilevel features within target regions. It utilizes object and focus attention modules to capture global and local features, and constructs channel attention coefficients to enhance channels containing important target features. (3) We propose a general, yet efficient dimensionality reduction method named adaptive PCA-whitening to improve short vector feature retrieval performance. This method can dynamically fuse the feature vectors of different dimensions via adaptive coefficients, yielding a more compact and robust representation. Extensive experiments on six benchmark datasets show that our method achieves better retrieval performance compared to existing state-of-the-art unsupervised methods. Compared with the next-best unsupervised methods, the mAP scores of our method (dim = 128) are 11.2 %, 6.1 %, 9.3 %, 8.3 %, 14.3 %, and 22.1 % higher on the Oxford5K, Paris6K, Oxford105K, Paris106K, ROxford, and RParis datasets, respectively.