Transductive Zero-Shot Hashing for Multilabel Image Retrieval

Locating Target Regions for Image Retrieval in an Unsupervised Manner

Image retrieval performance can be improved by training a convolutional neural network (CNN) model with annotated data to facilitate accurate localization of target regions. However, obtaining sufficiently annotated data is expensive and impractical in real settings. It is challenging to achieve accurate localization of target regions in an unsupervised manner. To address this problem, we propose a new unsupervised image retrieval method named unsupervised target region localization (UTRL) descriptors. It can precisely locate target regions without supervisory information or learning. Our method contains three highlights: 1) we propose a novel zero-label transfer learning method to address the problem of co-localization in target regions. This enhances the potential localization ability of pretrained CNN models through a zero-label data-driven approach; 2) we propose a multiscale attention accumulation method to accurately extract distinguishable target features. It distinguishes the importance of features by using local Gaussian weights; and 3) we propose a simple yet effective method to reduce vector dimensionality, named twice-PCA-whitening (TPW), which reduces the performance degradation caused by feature compression. Notably, TPW is a robust and general method that can be widely applied to image retrieval tasks to improve retrieval performance. This work also facilitates the development of image retrieval based on short vector features. Extensive experiments on six popular benchmark datasets demonstrate that our method achieves about 7% greater mean average precision (mAP) compared to existing state-of-the-art unsupervised methods.

Multi-Label Hashing for Dependency Relations Among Multiple Objectives

Learning hash functions have been widely applied for large-scale image retrieval. Existing methods usually use CNNs to process an entire image at once, which is efficient for single-label images but not for multi-label images. First, these methods cannot fully exploit independent features of different objects in one image, resulting in some small object features with important information being ignored. Second, the methods cannot capture different semantic information from dependency relations among objects. Third, the existing methods ignore the impacts of imbalance between hard and easy training pairs, resulting in suboptimal hash codes. To address these issues, we propose a novel deep hashing method, termed multi-label hashing for dependency relations among multiple objectives (DRMH). We first utilize an object detection network to extract object feature representations to avoid ignoring small object features and then fuse object visual features with position features and further capture dependency relations among objects using a self-attention mechanism. In addition, we design a weighted pairwise hash loss to solve the imbalance problem between hard and easy training pairs. Extensive experiments are conducted on multi-label datasets and zero-shot datasets, and the proposed DRMH outperforms many state-of-the-art hashing methods with respect to different evaluation metrics.