An Enhanced Deep Hashing Method for Large-Scale Image Retrieval

Hashing-based image retrieval plays an important role in approximate nearest neighbor search because of its storage and retrieval efficiency. Efficient image features are of vital importance for image retrieval task. However, the features of images may not directly suitable for image retrieval due to the illumination or cluttered background in images. In this paper, we propose an enhanced deep hashing method for image retrieval to promote the retrieval accuracy, in which we adopt an enhanced feature module that selects the attractive areas in images. It jointly explores the hashing learning, enhanced feature module, and batch normalization module in a unified framework, which can guarantee the optimal compatibility of hash coding and feature learning. The proposed deep model contains three parts: (1) a convolutional subnetwork consists of several convolutional-pooling layers and the proposed enhanced feature module; (2) a batch normalization module is utilized to control the quantization errors of hash codes at a moderate level; (3) a more comprehensive loss function is introduced to enhance the discriminative of image features and minimize the prediction errors of the learned hash codes. The experimental results on three datasets show that the proposed method can achieve competitive performance compared to other hashing approaches.