Research On Human-computer Interaction Question And Answering System Based On Image Comprehension

Due to the broad application prospect of image retrieval and query system based on image comprehension and the great application value,the research field of image semantic annotation is attracting more and more attention.The problem of image semantic auto-labeling is still a big challenge for researchers.If you want to mark the image semantically,you need to cross the huge "semantic gap" between the low-level visual features and high-level semantics.In order to solve the "semantic gap" between image features and image content,the researchers proposed a number of methods to mark the keywords,mainly around the two main methods of generating models and discriminant models.In this paper,a new method of image semantic annotation for fusing image category information.This method fully integrates the support vector machine to image classification and cross-media correlation model of the image tag probability relationship between the advantages of calculation,while complement each other.The image classification based on support vector machine makes up the problem that the cross-media correlation model is highly dependent on the clustering results.On the other hand,the image classification based on support vector machine utilizes the block color feature,and the cross-media correlation model is formed by dense SIFT The visual lexicon is used to avoid the problem of inaccurate labeling results caused by image segmentation.The candidate words produced by the fusion of the two are combined with the color feature and the regional characteristic information of the image,so that the accuracy and stability of the result Have improved.In order to further improve the accuracy of image annotation,this paper uses the above image automatic marking frame to obtain the candidate label words,the candidate label words for text mining.The purpose of using text mining is to find out the relationship between the text,through the relationship between the two candidates to determine whether the probability of a common description of the same image,so in the candidate marked word set the most likely set of marked as the final annotation result.In this paper,the Fp-Tree algorithm is used to extract the text of the candidate label words.In the process of application,the frequency of the marked word is merged with the probability of the standard word obtained from the improved model,and the probability and text The degree of annotation of the word parameters to determine the label phrase.Because the number of annotations of each image is not fixed,so the final result can not be given by way of interception,this paper takes into account the discretization of the probability of discretization,after processing,there will be marked between the classification of attributes,So we can determine the label words for each image.In this paper,the data is discretized by equal frequency discretization.Equal frequency discretization has the advantages of simple operation,small system time and space cost,and it is very suitable for the processing of a large number of candidate words.The experimental results were evaluated by image precision and recall rate and comprehensive evaluation coefficient(F).In this paper,the proposed method shows a very good performance compared with the traditional CMRM phase,which is based on the traditional image semantic self-annotation of translation model TM,cross-media correlation model CMRM,continuous space model CRM and so on.In this paper,the image semantic annotation method is improved by 24% and the precision of the image semantic annotation method is 12%.Compared with the CRM with good performance in image semantic annotation,Which increased by 14% and improved the precision by 6%.Compared with the traditional cross-media correlation model,the recall rate was improved by 18% and the precision was 20%.Compared with the current image semantics Compared with CRM,which has good performance,the recall rate is improved by 8% and the precision is 14%.It can be seen from the experimental results that the proposed method has improved the precision and recall rate,and the F measure has also been improved in the final result.It is not difficult to see that this paper has some application and research value in the field of image semantic annotation.