Research On Content-based Image Retrieval And Its Applications

In recent years, with the rapid development of modern science and technology, the scale of network image spresents an explosive growth tendency, especially when internet, social networks and other new technology are constantly emerging. How to solve the problem of using image resources reasonably, establish index efficiently, and retrieve the image of interest accurately in a large-scale database has become an urgent subject.This paper focuses on the research and applications of content-based image retrieval. The main contributions include the following parts:Firstly, the method of Dense-SIFT feature extractionis improved and a new adaptive extraction methodis proposed. Since the information of texture in each image area is not the same, using a fixed window will lead to either insufficient extraction in texture complex areas or overdone extraction in smooth areas. Therefore, we adjust the size of window adaptively based on the edge information of images to improve the description ability of the features, which can be represented by adopting K-means algorithm and Bag of Words(Bo W) model. Then, the feature vectors are constructed by combining the visual words histograms of Dense-SIFT features with the color features. On this basis, an image retrieval method based on adaptive Dense-SIFT(ada-Dense-SIFT)is proposed. The experimental results show that compared with the traditional Dense-SIFT feature, the proposed method can extract more image information and achieve better retrieval performance with 3.2% of the increase of precision.Secondly, in order to further improve the retrieval performance, a re-ranking method based on context similarity information is proposed. This method labels the correlation rating scores of the initial retrieved results automatically. We find the K-Nearest Neighbor(KNN) of each unlabeled sample from a labeled group, in which the k-nearest neighboris searched to construct the neighbor set. Then, the similar matrix is constructed by calculating the similarity between the unlabeled sample and each of their nearest neighbors based on two different methods. Next, the refactoring coefficient matrixis is calculated through the similar matrix and the final relevance degree can be obtained. After the above steps, the processing of re-rank is done. In addition, the re-ranking method based on context similarity information is compared with SVM-based and posterior probability based method aiming at improving retrieval performance and time consumption, etc. The experimental results show that the precision of the proposed re-ranking method can get 5.7% increase than that of the initial image retrieval method. Though the time cosumption of the proposed re-ranking methodbased on context similarity information is slightly higher than that of posterior probability based method, our method gets the best retrieval performance compared with the aforementioned two re-ranking methods.Thirdly, image retrieval is applied to the face retrieval area, and a retrieval schemeof face image for surveillance video is proposed. For the image from surveillance videos, the Ada Boost algorithm is introduced to judge whether the images contain human faces or not, andthe face regionis extracted from the face-existed image. Then, the block LBPH(Local Binary Pattern Histogram) feature is extracted with uniform partition method and local partition method. Next, according to the similarity, the search results are returned. The experimental results show that the proposed face images retrieval method can achieve a good applicability for low resolution surveillance video images which have complex background and various face pose. Meanwhile, the block LBPH with uniform partition method can achieve the best retrieval performance, under the condition of extensive consideration of the high and low resolution images.Finally, a face image retrieval system for surveillance video is realized. The system realizes the above techniques including face detection and partition, feature extraction and similarity measurement. And the proposed retrieval system can also achieve good retrieval performance when dealing with different resolution images.