Research On Image Annotation Methods Intergrating Multi-features With The Relevance And Diversity

With the continuous development of computer hardware and software, artificial intelligence technology, as well as a variety of intelligent electronic devices and the popularity of social networking sites and popularization.People has produced huge amounts of digital image in work, life, study, social and entertainment. How can in the Internet environment for rapid and effective image retrieval,becomes the hotspot in the field of artificial intelligence research. Because content-based image retrieval have the shortcomings such as slow processing speed, the effect is not ideal, the semantic of image annotation is a kind of mainstream solution. Its main methods include manual annotation, semi-automatic labeling and automatic labeling.Due to the high-level semantics is an abstract concept image, the same semantic corresponding image underlying characteristics differ in thousands ways, automatic labeling needs to solve the "semantic gap" between the two.In recent decades, in spite of all kinds of automatic image annotation methods are proposed,image automatic labeling are still facing spare sample,sensitive light, different scale, and many other challenging problems.Research paper is aimed at the problem, try to put forward a new solution, to improve the accuracy of image semantic annotation, and increase with the diversity of vocabulary.Basic ideas of the paper is as follows:appropriate model was proposed to solve the problem of high-level semantic expression and marked with the word synonymous.Before a problem mainly for annotation word corresponds to the image characteristics of the underlying difference is big, try to find stronger expression ability of probability model to describe with words of visual connotation;After a problem is mainly refers to different words of high-level semantics is consistent, is only a matter of expression is not the same. Try to mark vocabulary limited number of cases, find the difference semantic, and can accurately describe the image collection vocabulary at the same time.Based on the above research, the main content of the paper is as follows:(1) The paper in the third chapter mainly studies using mixture probability model to express the vocabulary from different angles of high-level semantics, and then complete the automatic image annotation. Specifically, first of all, to extract the image color, shape and texture characteristics, and then use each annotation words corresponding sample collection in color, shape, and texture three subspace training a Gauss model respectively, we call these Gauss model as annotation word sub-descriptor.Because the three feature subspace is the description of the different sides of the same things, to the image of the high-level semantic recognition ability is different, so you need to merge the Gauss models to enhance annotation vocabulary presentation skills.Three Gauss model with the method of weighted together constitute the annotation words description, and adopt the method of machine learning (paper based on the genetic algorithm) to learn the weights.Each annotation words training, for unknown image annotation process is as follows:the image segmentation, the segmentation of regional color, shape and texture feature extraction, and calculate its belong to the probability of each word, and the to select the annotation words for image region. The paper use Corellk test datasets of the above methods, the results of the experiment shows that this paper puts forward that the annotation word descriptor can effectively establish high-level image semantics and low-level vision features.The relationship between the labeling accuracy in average recall rate, average precision and average Fl measure is better than the CMRM value several aspects in major labeling algorithm.(2) In order to further improve the annotation of the Gauss mixture model is practical, the paper fourth chapter further study the correlation between fusion annotation words and enhance image annotation method annotation word diversity.Annotaion word's correlation refers to the correlation between image annotation words accompanying phenomenon, between some annotation words often is not independent of each other, but associated in common scenarios.Using its correlation can increase the accuracy of annotation.Annotation word's diversity refers to the vocabulary of synonyms or near-synonyms, they make annotation tag set redundant problems, especially with limited number of cases, the problem is more prominent.Paper mark word between the relevance and diversity, choose the largest amount of information, most likely for the image, and different annotation words of righteousness, further improve the performance of image annotation.Implementation method is as follows:first, using the words a symbiotic relationship and semantic relations associated to express the correlation between vocabulary, and use of the third chapter puts forward Gauss mixture model to calculate the diversity between different lexical semantics, and then Gauss mixture model is used to descriptor, the correlation between vocabulary and diversity to the given to mark images without annotation words.Among them, the first word annotation of selection using the model of the third chapter, namely, the calculation of image segmentation region corresponding to the maximum probability of annotation word descriptor select all area corresponding to the probability of the largest vocabulary as the first annotation, and select the candidate image annotation of words.Then annotaion word selection is calculated from the previous mark related words and can describe the image from even more angles of vocabulary.In order to verify the effectiveness of the proposed method, the paper selected Corel5k datasets experiment, the experimental results showed that this method with effect as of the third chapter, and increased with the diversity of vocabulary.The innovation of the paper lies in the following two points:(1) Put forward the fusion of multiple angle to express probability model to descriptor of vocabulary, using machine learning method to train the weight of descriptor.(2) Proposed a descriptor based on gaussian mixture model to calculate lexical semantic diversity,combined with the word symbiotic relationship, semantic relationship, the correlation between annotation and image regions, to choose high correlation and diverse annotation words.