Integration of Spatial Information Bag of Feature in Image Annotation
|School||Huazhong University of Science and Technology|
|Keywords||Target recognition Conditional Random Fields Contextual information Image Segmentation Graph cuts Energy Optimization Word bag model|
Image automatic target recognition marked an important problem in computer vision research , has been great development in recent years . The word bag model based on SIFT (Scale-invariant feature transform) the characteristics of the key points in the image scene classification and object recognition marked a good application . But there are still many problems , the scale transform light conversion , multi-angle changes , the difference between the same categories as well as the increase in the category of identification labeling caused great difficulties . For the traditional automatic target recognition marked by the presence of some of the problems in the following aspects have made ??some of the innovative work : 1) extracted based on a priori user interactive target framework to expand to two types of interactive image segmentation problem multi- class situation . The combination of multi-layered graph model , the use of the user's initial brush global energy function optimization , image segmentation results . 2 ) based on the dictionary space neighborhood characteristics , describing the image feature , taking into account the the word bag model ignores the structural relationship of the image space , the local spatial relationship of the image into a feature vector , improvement of image-based key point identification method , and this method is applied to a scene classification , achieved better recognition results . 3) integration of image pre-processing technology based on multi- partition and regional neighborhood histogram statistics , combined with the optimized conditions with the airport , to study the use of guide target identification problem of image segmentation , and then combined with top-down and bottom-up learning methods, to identify and locate the image of the target. Finally , by a plurality of classification vote decisions final result of the identification . Automatic image annotation systems and interactive object extraction system , given an image pattern recognition and machine learning methods , automatic annotation of the presence of the target and its location in the image , and reached more in some mainstream data set high recognition rate .