Scene Categorization Based On P.d.f Gradients And Feature Fusion

Scene categorization is a program that the image label is classified into predefinedcategories. Due to the complexity of image’s semantic information, the variation of scale,illumination, angle, and rotation, form distinctive image representation become animportant issue. In this paper, we focus on the methods of image representation andmulti-features fusion. We propose a scene categorization method that based on theorientation of p.d.f. Then we study on fusing multi-features to solve the problem of scenecategorization.Firstly, we analyzed the challenges which researchers are facing in the field of scenecategorization. That is how to better represent image’s content information to improve theaccuracy of classification. Meanwhile, the basic knowledge (thus as feature extractionmethod, kernel density estimation, SVM) to be used in this paper is also described.Secondly, Following the bag-of-features (BoF) approach, a plenty of local descriptorsare first extracted in an image and the proposed method is built upon the probabilitydensity function (p.d.f) formed by those descriptors. Then, we extract the features from thep.d.f by means of the gradients on the p.d.f. We construct the features by the histogram ofthe oriented p.d.f gradients via orientation coding followed by aggregation of theorientation codes. On this basis, a method of image categorization based on histograms oforiented p.d.f gradients is proposed. In this method, Support Vector Machine is used forconstructing a multi-class linear classifier.Thirdly, because of different feature extraction methods focus on different imagecharacteristics. We aggregate several existed feature and our p.d.f gradient feature to makefull use of various features. Then the label of scene image is determined by them.Finally, we codes the proposed methods using Matlab tool, and analysis its results toprove its feasibility and effectiveness.