Research On Multi-features Fusion For Indoor Scene Classification

Indoor scene classification is a research domain of scene classification, which is also a research challenge of scene classification. If solved the problem of indoor scene classification in low, which will help us improve the classification accuracy of the whole scene, at the same time, more effectively applied to scene images in areas such as video retrieval and robots. So there is important research significance for improving the accuracy of indoor scene classification.In view of the current applications of indoor scene classification of two feature models:one of models is that the pyramid model is put forward by David G and the pyramid model of the gradient histogram is put forward by Bosch A; Other is that Oliva put forward by integral space model of the envelope. This paper proposes the indoor scene classification method which is based on the fusion of global features and local features. This method combines the detail of the local features and the whole picture of the global features. It effectively improved the indoor scene classification accuracy. So this research has the very high research value.Firstly, this paper summarizes the related background knowledge of the indoor scene classification and its research significance; the indoor scene classification research status was analyzed.Secondly, this paper introduces the detail proceeds to obtain SIFT, Gist and PHOG features. Also, it introduced SVM and PCA. The SIFT and PHOG features belongs to the local features of indoor scene images and the Gist features belongs to the global features of indoor scene images. In addition, this paper introduced a multi-features fusion’s method, which is a very important role for how to determine the weights in each feature.Thirdly, the improved SIFT feature extraction method is proposed in this paper. In the scene image SIFT feature extracting classic space pyramid model generates n*128characteristic matrix, where n is the numherof key points, namely the matrix rows,128for the characteristics of the key dimensions. In this case,a large number of zero elements exist in image characteristic matrix, resulting in scene classification accuracy is very low. Therefore, clustering based on the key position and then extracting the SIFT features obtain unified dimension of SIFT feature matrix, then reducing dimensions. In addition, the specific steps of the Gist and PHOG features were extracted were expounded. Also, the multi-features fusion of indoor scene classification’s method was specifically analyzed in overall architecture and the generation of indoor scene multi-features.Fourthly, in this paper, the experimental results are analyzed. In the first, the experimental proportion is in the training set and testing set of the indoor scene images. The best proportion obtained from the experimental results, which help the indoor scene classification. Then, proceed the indoor scene classification experiment. By comparing the classification shows that compared with the single feature indoor scene, multi-features of the indoor scene classification did have advantages. At the last, on the basis of the first two experiments, improve the indoor scene classification of multi-features fusion experiments and further improve the accuracy of indoor scene classification.Finally, Summarize the general characteristics of the indoor scene classification. for the current research contents were also discussed at the same time, put forward the need to improve place for further research.