Deep Learning Based Hashing Method For Multimedia Retrieval

With the rapid development of storage and communication technology,the amount of information on the Internet explodes,especially multimedia data.This triggers the change of image retrieval,from text-based retrieval to content based retrieval.In order to improve the efficiency and accuracy of image retrieval,many researchers proposed multiple models to deal with retrieval problem.Deep hashing based method is the hottest method with best accuracy.'Hashing' means that all the images or videos are mapped to binary codes during the retrieval process.And hamming distance is the way to evaluate the similarity.'Deep' means that the hash function is learned with deep network like convolutional neural network.The existing hashing method can not generate hash code with large hamming distance.The accuracy of the retrieval problem can be further improved.Our paper deals with this problem,we propose a deep learning based hashing system for image retrieval.The key parts of our method is target hash code generation algorithm and two deep network for input images of different sizes.Experiments show that our method has leading performance.We further expand our proposed system to video retrieval and do experiments to prove its effectiveness.We use image retrieval as an example to describe the background and current research status in this paper.We first describe the definition,evolution and high-demand of image retrieval.We show the pipeline of normal image retrieval and it lead to the more efficient hashing method of image retrieval.Then we describe the categories of different hashing methods.The success of deep learning in computer vision also triggers the deep hashing method for image retrieval.We introduce the details of convolutional neural network and give a brief introduction of the deep hashing method.Deep hashing method can be further improved though it is the current best method for image retrieval.We propose a deep learning based hashing method for image retrieval.Our method first generates a target hash code set.This code set satisfies that each pair of code words in it have large hamming distance.It can hold the semantic information of the image labels.We use the target hash code set to generate new form of training data and realize the point wise hash learning.Our hash learning network is based on convolutional neural network and the last layer of it is a hash layer which can generate binary codes.For different input size,we design two networks.These can be used based on the dataset and the need.We perform image retrieval experiments on three datasets MNIST,CIFAR-10 and ImageNet.Our method is compared to more than ten hashing methods.And our method reaches the best performance in retrieval's mean average precision.For ImageNet dataset,we improve the MAP by nearly 10% comparing to other methods.The result proves that our method is very effective.Finally,we expand our system to suit the more complex video retrieval.We added the key frame extraction part and video hash code generation part to the system.We perform experiment on UCF-101 dataset,the MAP is larger than 0.9.It proves that our retrieval system can be expanded to video retrieval and the performance is still good.