Classification And Recognition Of Image Based On Local Features And Weakly Labeled Data

In order to realize the similar textual expression of image content,the key problem is to establish the relationship between them.The probability theory provides a solid mathematical foundation for solving this uncertainty reasoning problem.In recent years,the probability theory has been used widely through the complexity of inference calculation has been rapidly reduced.This led development of the applications of computer vision which are based on probability graph models,such as image segmentation,motion detection and tracking,target recognition,and scene understanding that requires a comprehensive consideration of various factors.In this paper,we use existing methods of image classification and recognition to develop how to extract effective image features to achieve accurate image classification.For the lack of annotation on pixel-level image,it is necessary to consider adding certain constraint information to relative methods about the semantic information of images.We model the joint probability topic hierarchical structure for solving the problem of similar object discovering and recognition under the same category.In addition,we also afford a thought of solving the problem of scene understanding through improving topic model with given objects semantic labels.Our key contributions include:(1)The support vector machine(SVM)is insensitivity for missing data and cannot afford general solution to nonlinear problems.Starting from the idea of simplifing complex problems,we split the image dataset into several subsets for computering.To avoid the uncertainty of kernel function selection and the nonlinear data falling into the local minimum value problem,we introduce histogram kernel function instead of the traditional Gaussian kernel method.We also demonstrate that using linear kernel function is better for mapping feature to high dimensional space to realize fast and accurate classification on small sample data.(2)A soft-partitioning spatial pyramid matching method(sSPM)is proposed to solve the problem of semantic ambiguity of boundary features,which are caused by pyramid hard partitioning.Research on the allocations of local feature along the boundary in the spatial pyramid,we design a candidate region to hold them and construct distance function between the features in candidate region and the adjacent image block.The robust feature vectors are computed through judging the similarity between the features in the above two regions.In addition,Lagrangian multiplier and Strong dual principle method are used to design an optimized multi-kernel classifier.Comparing with other similar methods in the experimental section,the results demonstrate that the proposed method “sSPM+MKL” is effective and efficient in image classification.(3)Owing to the image differences between the intra-class and inter-class are not obvious,the object co-localization method of sharing the same weakly semantic information is proposed.Research on the corelations of shared part feature and weakly annotations,we establish a shared pool for "object-feature" parts to bridge the "feature-part" and “part-object” correspondence relationship.In addition,research on the composition of object with different proportions calling shared parts,we realize similar object detection under weakly supervised learning method.In order to eliminate the interference of the noise image,the information entropy is used to measure the similarity of the images.Defining those images which don't transfer information in the sharing pool are noise images.The results demonstrate that the object discovering and localization without interference can sufficiently improve the IOU performance.(4)Due to the lack of pixel-level annotation of image samples,we propose a semantic "Context+Focus" method for scene segmentation.A large number of image weakly tags and features are used to establish "top-down" and "bottom-up" mechanism to learn the relationship between semantic object and feature object.A hierarchical probability graphic model is modeling to correlate the correspondence relationship of tags and objects.The "focus" means the key object and "context" means semantic tags of all the objects in the scene.By the means of Blocked-gibbs algorithm and maximum a posterior,we form a multi-region-multi-object semantic linkage segmentation to infer the semantic labels of each region and object in its region.