Research And Realization Of Image Annotation Algorithm Based On Image Feature And Context

In recent years, with the development of information technology, Internet and especially the development of mobile internet, a variety of information data on the Internet explosion, people are steping into the era of data from the information ear, arrangement, analysis and understanding of this data will provide tremendous value. Unlike traditional structured data, most data on the Internet are unstructured data and many of these unstructured data are images, so it becomes more and more important to analyse and research these images on the Internet. Due to the exists of "Semantic Gap" between the people’s understanding of semantic information and the features of images, it is difficult for users to find out the interesting images they want from the massive Internet pictures, people urgently need a method to find the desired pictures facilitatly. So, how to connect the high-level semantic and low-level image features is becoming a research focus. The Automatic Image Annotation(AIA) is a good method to fill the gap between users and pictures.Traditional Automatic Image Annotation method only focus on the features of images and ignore other information, but in the ear of Internet, most image stored in the Internet, these images have rich contextual information, the analysis of such information will help to produce more accurate and abundant annotation of these images.This thesis research the problem of automatic image annotation and come up with a method which combine feature and context of the images to automatic annotate images, there are two steps of this method, first, we annotate the images through features; second, we use the context information as supplement annotation.We use feature extracted from deep learning and SURF algorithm at the same time when we annotate the images, then we combine the weighted annotation result from each feature. Unlike the traditional method which train a magical classifier to classify the images, then use the classify result to annotate images, this thesis use a new method, this method use the similarity between unannotated images and annotated images to annotate the unannotated images. Then we transfer the problem of images annotation to images retrieval.In order to speed up the image retrieval process, this thesis use the B+trees’forest structure to realize the fast Approximate Nearest Neighbor search, this structure divide the sample images’ feature through random plane, features in different part of the plane will be parted into different subtree of the random plane, then use different random plane to divide the features in the subtree, repeat this process until the num of the features is under the threshold. Then we get a structure of B + tree, many of this kind of B + trees make up a B + trees’ forest.When an image need to be annotated, we use the image’s feature to search in the B + trees’ forest and get the images whose features are close to the the need annotated image’s feature. We use the weighted summation of the search result’s annotation to get the image’s deep learning annotation.As for the SURF feature’s annotation process,.the principle are just like the deep learning annotation process, we search the TopN most similar images and then use the weighted summation of the search result’s annotation as the annotation of the image which need to be annotated. Then the image’s feature annotation can be get through weighted summation of deep learning annotation and SURF annotation.The context of the image always contain information which is related to the content of the image, we can get more accurate annotation if we mining the information of image context. In this thesis, we use the context information of the image as supplement of the image annotation. We use RAKE algorithm to extract the keyword of the context and get the keywords with their weighting factor, to avoid introducing irrelevant information to the finally annotation, we weighting the context annotation with the Jaccard distance between the feature annotation and context annotation. Then we combine the weighted context annotation and feature annotation and sort them by there weight, pick the annotation whose weight is bigger than the threshold then we can get the final annotation.Compare with the traditional method, the advantage of the method in this thesis is that we need not to train each annotation and its high efficiency, because of the annotation of this method are not limit, the annotation can be more rich.