Research Of Image Annotation Based On Image Segmentation And Regional Semantic Correlation

With the rapid development of computer technology, network technology and intelligent communication technology, a large number of image data is widely spread on the network, and showing explosive growth, how to effectively manage and use these image resources has become a difficult problem.Although people has made a lot of results in the field of image retrieval,there are still many problems.Text-based image retrieval has long been unable to meet the needs of the current large data age due to the low efficiency and human subjectivity.Content-based image retrieval can not solve the"semantic gap"problem which hindered its development.Semantic-based automatic image annotation is the main direction of the development in image retrieval filed.Although researchers has done a lot of research and exploration in the field,there is still a lot of difficult problems.Aiming at the research status and development trend of the field of image retrieval and the problems that we are facing at present,this paper proposes a series of effective methods to improve the performance mainly in the following ways:(1)Semantic based image annotation requires the use of image segmentation algorithm for image pre-processing,and an accurate and effective image segmentation is very important for image feature extraction and annotation model construction.This paper proposes an improved image segmentation algorithm,the basic idea of the algorithm is:Firstly,we use Mean Shift algorithm for image pre-segmentation.Due to Mean Shift algorithm is sensitive to the image edge,it is very good to extract the edge of the image,but it is also very easy to produce a lot of small areas. In view of this shortcoming,we use the Ncut algorithm to further deal with the image area obtained in the previous step. Since the Ncut algorithm always tends to get a larger image area, it can solve the over-segmented problem of Mean Shift.Since Ncut handles the image area that has been segmented, rather than the pixel, it greatly reduces the amount of computation and improves the performance of the algorithm.However, the Ncut algorithm also has some shortcomings, the algorithm is a NP-hard problem.Before dividing,it needs to specify the number of areas you want to split first.If the parameter is set incorrectly,it is easy to get over-segmentation and under-segmentation.Therefore,we use the region merging and splitting algorithm to further correct the segmentation result obtained by Ncut processing.merge the over-segmented regions and split the under-segmented regions, so that improve the accuracy of the image segmentation results as much as possible.(2)In this paper,we propose an improved image semantic annotation method combining with region semantic correlation and Gaussian mixture model.In the traditional Gaussian mixture model, it gets the image annotation results directly based on the semantic posterior probability :one is to get the semantic words of the largest n semantic posterior probability as the result of the image annotation , the other is to directly select the semantic words whose semantic posterior probability is larger than a certain threshold as the result of the image annotation . And the method above can not get a accurate result of the image annotation, it is easy to get some redundant and incorrect semantic words, which influences the accuracy of the image annotation results.Moreover, considering the "semantic gap" problem in the model, the posterior probability does not completely determine its weight, and there may be a large error only based on the posterior probability. In view of the above problems, we proposes a GMM image annotation method based on regional semantic correlation, which integrates the regional semantic correlation into the GMM model to get a comprehensive result,which effectively calibrate and optimize the annotation result of the model, so as to improve the accuracy of the annotation results.