Research On Image Retrieval Technology For Specific Object

In recent years,with the fast development of multimedia technology,the media data such as image and video on the Internet increase rapidly.Image as a mainly form of media data,which can deliver a wealth of information.Hence,how to retrieval similar images from such a large-scale dataset becomes a tough problem in multimedia field.The framework of content-based image retrieval usually includes database building stage and query stage.In the database building stage,firstly,extract the visual features of all images in the database;secondly,organize the features to build index.In the query stage,firstly,extract the same feature for query image;secondly,search in the high-dimension index;finally,return the similar images.In many practical applications,the purpose of users is not only to retrieval images that are similar to the query image as a whole.They prefer to get the images that contain the specific objects,namely image retrieval technology for specific object.This technology has a great study value on many applications such as medicine,transportation and so on.However,the problems of the general image retrieval framework are as follows:the specific objects tend to be small in images,and there is a wide variety of objects.It is hard to distinguish them from the images with complex background.These attributes increase the difficulty of detecting and recognition.Hence,in order to achieve fast and accurate image retrieval for specific object,this paper conducts research on image retrieval technology for specific object.The main contributions of this paper are as follows:(1)Detection for specific object based on convolutional neural networkTo handle the problem that the area of the specific object in the image is usually small,this paper proposes a specific object detection method.The method optimizes Faster R-CNN for small objects,and utilizes two region proposal networks to fuse the feature of the middle layer and high layer and obtain more candidate regions.In addition,the feature map of the high layer is enlarged to increase the accuracy of locating objects.The experiments are conducted on LUNA 16 which is a public pulmonary nodule detection dataset.The recall of the proposed method can reach to 86.42%,which increases 3.3%compared with other existing methods.The results demonstrate that the proposed method is effective for detecting small and distinguishing difficultly specific object.(2)Hash coding based on convolutional neural networkDue to the fact that the image data in the Internet is very large,it will take up a lot of storage space if images are represented by real-valued features and much time will be cost for similarity computing.This paper proposes a hash coding method based on convolutional neural network,which employs the similarity relationship between pairwise images,and takes advantage of quantization loss minimized and information entropy theory to increase the mean average precision by 5.98%compared with existing hashing method,which only uses 12 bits.(3)Image retrieval system based on binary indexThis paper designs an image retrieval system based on error weighted hashing to further improve the speed of the nearest neighbor search.The main functions of the image retrieval system:firstly,the binary codes of images are extracted by the proposed hashing method;secondly,construct binary index table;thirdly,for query image,implement k-nearest neighbor search in hamming space to screen candidates out based on error weighted hashing;fourthly,screen the results which contain the specific objects.The experiments on billion-level database show that the precision can achieve 96.8%,and the speed is 0.31s per image which 82ms faster than linear search.