Research Of Image Retrieval Based On Saliency Detection And Hashing

With the development of science and technology,image data is growing exponentially.It takes a lot of time to retrieve the corresponding picture from the database.Content-based Image retrieval technology can be roughly divided into two categories: whole-image-based image retrieval algorithms and salient-regions-based image retrieval algorithms.Whole-image-based image retrieval algorithms needs to extract all the features of the whole image,which often contains a lot of background noise.The salient-regions-based image retrieval algorithm is better than the whole-image-based image retrieval algorithm because it takes the semantic information of the image into consideration.Hashing method,which converts image data into binary code,reduces the storage space of image data and improves the retrieval speed,and is increasingly applied in scientific research.Based on the above characteristics,we propose a image retrieval algorithm based on saliency detection and hashing.The main work of this paper is as follows:First,a saliency detection method based on multi-image fusion and multi-feature is proposed.Traditional saliency detection algorithms based on manifold usually construct a single graph to describe the relationship between different regions of an image.However,images of natural scenes often have complex structures between image regions,using only one image may ignore the important information of the image.This paper proposes an algorithm that constructs multiple graphs to describe the image information with different feature spaces.Specifically,two graphs are constructed based on spatial location and color feature respectively: one is a K regular graph based on spatial location,the other is an ? graph based on color feature.Secondly,the rarity term is introduced into the saliency optimization framework.The traditional saliency detection framework based on manifold is constructed by using the information between connected nodes and the information of query points in the smoothness constraints.The cognitive property of visual is not included in this framework.Therefore,the performance of this framework is limited when it is applied to object detection.The algorithm in this paper focuses on the graph-based optimization problem itself and introduces a new graph-based optimization framework to overcome the above problems.Thirdly,a strongly constrained manifold hashing(SCDMH)is proposed for each length hash bit.The mutual reconstruction term between the original feature and the hash codes is added into the objective function.The information loss can be reduced by minimizing the reconstruction loss.At the same time,this paper embeds manifolds into supervised discrete hashing for the first time to learn and preserve manifold structures directly in Hamming space.Fourthly,an image retrieval system is designed and implemented by combining the improved saliency detection and hashing algorithm.Three experiments have been designed in this paper.About saliency detection,the method proposed in this paper showed good performance on PR curve,F value,AUC and other indicators in three public datasets.For strongly constrained manifold hashing,compared with five unsupervised hashing methods and four supervised hashing methods,the hashing method proposed in this paper showed good retrieval performance on all indicators in CIFAR10,Caltech-256 and MNIST.Finally,the paper combined the two improved algorithms,and achieved high accuracy in the image retrieval system.