Feature Extraction And Semantic Annotation Research On Criminal Scene Investigation Image

In the field of the current domestic public security,it is basically relying on the way of pure manual annotation for the criminal scene investigation(CSI)images,which is not only inefficient but also subjective.In order to solve this problem,this article puts forward some improvement methods about image feature extraction and semantic annotation on the basis of reading a large number of references at home and abroad,which may be helpful for realizing automatic semantic annotation of the CSI images and reducing the work intensity of the public security police.Firstly,we made further studies on color and texture features of the image,which are both low-level features.On the basis of these characteristics,a fusion feature extraction algorithm was proposed by adding different weight coefficients.Through lots of experiments,we regarded the best weights in the experimental results as the parameters of the fusion feature extraction algorithm.Based on the thought of image classification,we considered the image semantic annotation as an image classification problem.Through the way of feature histogram similarity comparison we realized image classification,and then the CSI image semantic annotation would be implemented.Experimental results show that the image semantic annotation accuracy of fusion feature algorithm is greatly improved than the two separate algorithms.Secondly,we studied the characteristics of SIFT and SURF,in view of the SURF feature is not an ideal situation in image scale and rotation changes performance,we took the thought of SIFT feature and on the basis of SURF feature proposed the GP-SURF feature extraction algorithm combined with the Gaussian pyramid model.The core idea of the algorihm is in the stage of constructing scale space adopting Gaussian difference pyramid in which the image size is changeable,which can simulate human being;s watching from far and near while keeping the scale invariance.In this way we expect to overcome the influence of image scale and rotation changes and to improve the characteristic representation for the image.The experimental results show that the GP-SURF feature significantly improves the semantic annotation accuracy of CSI images.Thirdly,we made further research on the Bag of Words model.On the basis of this model,we clustered extracted image features through a clustering algorithm,and the results were regarded as visual words,all of which made up the visual dictionary.Then through training the Support Vector Machine,there would be lots of classification hyper planes for all kinds of CSI images.Next by pair wise comparison the Support Vector Machine classifier was constructed,through which we could realize classification and labeling.And then a model of the image semantic annotation was implemented.Through the actual experiment process,the effectiveness of the proposed image semantic annotation model is verified.Finally,we implemented a "Legal Case&Image Management and Retrieval System".Through putting forward the fusion of color&texture features and GP-SURF features again,we got the fusion feature called HL-GS.Based on this characteristic,we implemented the semantic annotation of CSI image through a fusion feature semantic tagging model,which had been applied to the semantic tagging module and image retrieval module of "Legal Case&Image Management and Retrieval System.