Research On Automatic Image Annotation Method Based On Dictionary Learning

With the rapid development of computer technology and mobile Internet,a large amount of image information is generated in different ways every day and uploaded to the Internet in the form of numbers every day.This makes images become an important means for people to transmit information in the current Internet age.Faced with an ever-increasing number of image resources,it is extremely urgent about how to quickly and easily organize and accurately retrieve a target image.Automatic image annotation is a research hotspot in the current image processing field.Its purpose is to assign a number of tags(keywords)to the image that can describe the visual content of the image.It is a key step in the current text-based image retrieval process.During the past ten years,Automatic image annotation has achieved fruitful research results.Not only the accuracy of the labeling,but also the marking model has been continuously improved.However,because of the semantic gap between the low level visual features and the high-level semantics,the annotation performance and the labeling efficiency of the automatic image annotation still need to be improved.In this paper,based on the study of existing annotation methods,starting with improved labeling performance and labeling efficiency,combined with sparse coding techniques in machine learning and the efficiency of dictionary learning methods,this paper proposes an automatic image annotation method based on Fisher dictionary learning and an automatic image annotation method based on discrimination dictionary learning.The specific work is as follows:(1)First,it briefly introduces the research status of automatic image tagging,and the difficulties and deficiencies in current research.Then it introduces the basic theoretical knowledge related to the process of automatic image annotation.(2)In order to reduce the impact of semantic gap on automatic labeling of images and improve the labeling efficiency,this paper proposes a method for automatic image labeling of multi-label discrimination dictionary based on the high efficiency of label consistency discrimination dictionary.The method firstly extracts multiple types of features for each image in the input feature space of the dictionary learning,then introduces a regularization item related to the test set label,and uses the tag correlation between samples as part of the input character data for dictionary learning.Finally,based on the discriminant dictionary and coefficient matrix obtained,the corresponding label prediction algorithm is designed to realize the semantic annotation of unknown images.The final experimental data shows that using this method can effectively improve the accuracy of labeling and labeling performance.(3)Aiming at the problem of unbalanced distribution of dataset labels and latitude of effective combination of different types of image features,an automatic image annotation method based on the Fisher discriminant dictionary is proposed.This method maps the initial input data to a high-dimensional kernel space for discrimination dictionary learning under the action of a Gaussian kernel function.Experimental results show that the automatic image annotation method based on nuclear space Fisher discriminant dictionary learning has certain effectiveness and feasibility in improving image annotation performance.