Research On Algorithm Of Entropy In Image Retrieval

With the vigorous development of Internet technology,various kinds of social software come into our lives.Every day,we need to contact various forms of multimedia information.How to extract the information accurately and make effective use of it has become an important research topic,which has attracted wide attention of the academic community.In this context,image retrieval technology has been fully and comprehensively developed.Early image retrieval was mainly text-based.There are two main problems in this method: Firstly,the text information of the image depends on manual labeling,which requires a lot of time and money;Secondly,for some images with complex background,manual labeling can't accurately express the main information of the image.Therefore,content-based image retrieval quickly replaces text-based image retrieval.Content-based image retrieval no longer relies on manual annotation.It automatically analyses the main content of the image and generates image descriptors through the algorithm.Content-based image retrieval has two main stages of development.Initially,content-based image retrieval mainly relied on manual design features such as BoW,VLAD and Fisher Vector.After the processing of Gauss convolution,local feature aggregation and normalization,these features have good stability and they are easy to calculate.At present,due to the great improvement of GPU computing power,image retrieval algorithm based on deep convolution network has become the mainstream.With the support of hardware,the number of layers of convolution network is increasing,and the types of network are becoming more and more diverse.Whether for traditional image retrieval algorithms or image retrieval algorithms based on deep convolution network,the similarity of descriptors is measured by calculating the Euclidean distance between descriptors.In this paper,we find that it is not enough to describe the differences between features only by Euclidean distance.For a set of matching feature points,the similarity will be destroyed by changing the distribution state of the feature,but the Euclidean distance can't get the change.Based on this theory,the entropy boosted loss function and spatial distribution entropy are proposed in this paper.Spatial distribution entropy adds the distribution state of descriptor to image descriptor.Entropy boosted loss function emphasizes the difference of distribution state between features in the process of network training.In summary,the following research results have been achieved in this paper:1.In order to improve the accuracy of image descriptors,this paper presents a method of describing the spatial distribution of local features in images by using spatial distribution entropy.Each image counts the spatial distribution of local features from three aspects: spatial coordinates,scales and directions,finally generates frequency distribution histogram and calculates spatial distribution entropy.After normalization,spatial distribution entropy is joined with image descriptor,and an improved image descriptor of spatial distribution entropy is obtained.Spatial distribution entropy can be applied not only to traditional image descriptors but also to image descriptors of deep convolution networks.In this paper,the effect of spatial distribution entropy is proved by many experiments such as image retrieval,image classification and massive image retrieval.2.Entropy boosted loss function is presented in this paper.At present,deep convolution network has made remarkable achievements in the field of image retrieval.In order to further improve the accuracy,image descriptors need to capture some subtle differences in highly similar images.At the same time image descriptors must easily distinguish images that do not belong to the same category.The entropy boosted loss function proposed in this paper retains the part of the comparison loss function that calculates the Euclidean distance between images.At the same time,the new loss function calculates the difference of distribution state between image features.Finally,the entropy boosted loss is obtained by adding the two parts together.The experimental results show that the entropy boosted loss function improves the accuracy of image retrieval.