Attention-Based Video Hashing for Large-Scale Video Retrieval

Abstract

Large-scale video retrieval is a challenging problem because of the exponential growth of video collections on the Internet. To address this challenge, we propose an attention-based video hashing (AVH) method for large-scale video retrieval. Unlike most of the existing video hashing methods, which consider different frames within a video separately for hash learning, we use a convolutional neural network and long short-term memory (LSTM) network as the backbone to learn compact and discriminative hash codes by exploiting the structural information among different frames. To better capture informative clues in the video, an attention mechanism is added into the backbone, which can assign different weights to different LSTM time steps. Experiments were conducted to evaluate the proposed AVH method in comparison with existing methods. The experimental results on two widely used data sets show that our method outperforms existing state-of-the-art methods.