Image Scene Classification Based On Feature Combination

The image scene classification is an important part of the high-level semantic image understanding, aiming at the semantic annotation process by analyzing the image on the overall statistics and associated features. Scene classification belonging to the major branch of computer vision has an important and widely used in variety field. The automation of scene classification arises at the historic moment due to the conventional method that relying on artificial classification obviously can not meet the demand.Many research of scene classification algorithm have been done in recent years, but the algorithm have defects in different degrees, such as the Gist feature can not describe the local information of image very well and the local Gist feature is difficult in calculation etc. First, this paper proposes a method that based on improved Gist feature to extracting the image features. Second, combines with the Gist feature and the PHOG feature to implementing the scene classification. Finally, designs the system of scene classification. The system can be well applied to scene classification. According to the different requirements of scene classification, the mainly works are as follow:1. First, this paper introduces the characteristics of Gist; second, tests the different specification of the grid that is used to divide the scene image; third, filters the image with filter bank; finally, cascades the value to getting the description of image features. In order to improve the accuracy of scene classification, a mesh refinement of the scene image is used and the most suitable grid image is obtained by experiment testing in this paper, which offers the explanation of the appropriate grid mesh generation in the subsequent scene classification system.2. The improved Gist feature extraction is introduced in detail. In order to make up for the deficiencies of Gist features and take advantage of PHOG features, a combination of Gist features and PHOG features is proposed, which give a better description of the image scene information. Based on this combination vector, scene classification model and scene classification test are gained through the training classifier in machine learning.3. This paper divides scene into natural scene and artificial scene, then using the existing second classifier to classify it. Compared with several existing methods of the scene classification, the experimental results show that the accuracy of this method can effectively improve the scene classification. For natural scenes and artificial scene classification accuracy can be improved to 96.84%. This paper designs a multi-stage classification system in order to distinct previous one-to-one or one-to-more classification model in multiple class classification system. First, the system divides scene into natural scene and artificial scene, then uses a small category of natural scene and artificial scene to training the different classifiers, finally combines a scene classifications system according to the different settings and completes the overall design of scene classifications system. Compared with the existing centralized classification method, the result shows that the proposed classification system has certain improvement in classification time and classification accuracy.Through the artificial-natural scene classification and multi class scene classification comparison and analysis of test results show that the method is feasible.