Triplet-Based Deep Video Hashing For Large-scale Similar Video Search

With the rapid growth of the equipment and bandwidth of the Internet,uploading and downloading images and videos on Internet has become more convenient.This leads to a tremendous growth of the video data,and raise the challenge of storing and processing.How to search and retrieve videos that users required in massive video data accurately has become a real application problem that needs to be urgently solved in the multimedia field.Hashing attempts to convert high-dimensional original data into sets of short binary codes,and preserve the similarity of data points in original data space.Hashing can accelerate the retrieval speed and reduce the requirement of storage.Therefore,hashing based image and video retrieval has been a research hot spot in recent years.Hashing has been widely studied in image retrieval,however,video hashing has attracted less concerns,the main reasons are video hashing usually lack the ability of feature representation and label embedding,and has high complexity of computation,this leads to a large performance gap compared to image hashing methods.In this thesis,we choose the learning-based video hashing as our research topic to improve the performance of video retrieval.In this thesis,we propose a triplet-based deep video hashing method,and conduct experiment to evaluate the effectiveness of our method.The main work of this thesis are described as follows:(1)In this paper,we propose an end-to-end supervised video hashing network,the network aggregates the convolutional neural network and long short-term memory.We extract high-level feature in frames by convolutional neural network and fuses the temporal information in long short-term memory,and effectively extract video-level features.(2)Our model is trained by triplet similarity labels and adds a classification model to reinforce the supervise information embedding to make model more discriminative.Besides,the loss function aggregates the embedding of triplet similarity loss and classification loss,which can effectively generate the more discriminative hash codes and preserve the data similarity in original data space.(3)The proposed method adds additional constraints to binary hash codes in deep neural networks.The binary constraints improve the quality of hash codes,and beneficial for hash retrieval.These constraints can not only make hash codes more balance,and reduce the correlation between different hash bits,which increases the information entropy and reduces the information redundancy in hash codes.The performance on three large-scale video dataset illustrate that proposed method improve the performance of video retrieval,and outperforms existing hashing approaches on mean average precision,precision-recall curves and several other evaluation methods.