| With the development of information technology,the image data on the network shows explosive growth,and how to quickly and accurately retrieve the images that people are interested in from the massive image data has become an urgent problem to be solved.The image retrieval technology arises at the historic moment and become a research hotspot in the field of computer vision.On the basis of analyzing and summarizing the existing image retrieval technology,we put forward a more effective image retrieval scheme,and implement an image retrieval algorithm which can meet various practical needs of users.The current image search engines usually extract the whole image features to build an index to complete the retrieval task.However,in many cases users focus on only a part of the image,i.e.object-of-interest.As a result,the features extracted from the image are partially effective.In other words,some of the features are ineffective and could have a negative impact on the retrieval process.To overcome this difficulty,an image retrieval scheme based on object-of-interest is proposed,and collaborating with the existing techniques in saliency detection,image segmentation,and feature extraction,an effective image retrieval algorithm is coded.In the first part of this paper,the hierarchical saliency(HS)detection algorithm is firstly adopted to analyze the user’s object-of-interest,and the saliency-based image cut(SC)algorithm is employed to segment it.Then,we extract the HSV color features,SIFT local features and CNN semantic features of the object-of-interest.Finally,the similarity between the object-of-interest and every database image is computed and the retrieval result is sorted accordingly.In the second part,combined with the current mainstream CNN models,a method of using fused CNN features to express the image content is proposed,based which an image retrieval algorithm with higher performance index is completed.Specifically,three different kinds of CNN features are extracted from the pre-trained models,respectively.On this basis,the weighted average of the similarity scores of the query image and database image is calculated,and images with high similarity score are returned to the user as the retrieval result.Extensive experiments on two publicly available datasets well demonstrate that the image retrieval algorithm based on fused CNN features is clearly better than the retrieval algorithms based on individual CNN features and other current image retrieval algorithms.On the basis of analyzing former two parts,we define the interested objects more broadly in the third part of this paper.The background and salient objects of the image can be regarded as the user’s interested objects.After a full study and analysis of the user’s needs to retrieval images in a variety of situations,a guided retrieval scheme is designed and the corresponding algorithm is implemented.Firstly,a new strategy is proposed to solve the problem that multiple salient objects may appear in the image,which calls the SC algorithm in an iterative way and segments the multiple salient objects sequentially.Then,this multiple salient object cut strategy is used to understand and obtain the user’s interested objects.Before performing the actual retrieval,the algorithm will provide users with a variety of retrieval situations based on different interested objects in the image,and guide the user to perform effective image retrieval.Simulation experimental results show that the guided image retrieval algorithm is clearly higher than other retrieval algorithms in the performance index,and in the actual implementation of retrieval,our algorithm can largely solve various image retrieval tasks of users. |
