Reseach On Scene Classification Algorithm Based On Local Feature Extraction

With the rapid development and popularization of the Internet, the amount of Internet image data grows exponentially. Facing such a lot of image resources, how to supervise and manage them quickly and efficiently has become a hot research. And the important way to solve such problem is scene classification technique. Scene classification aims at annonating massive image data into different classes automatically, then make automatic image classification come true. It has great research value for overcoming problems, such as heavy workload, low efficiency, high cost, fatigue and easy to misjudge, during the annonating by people. The key problem in scene classification is how to use low-level feature to obtain category information, and local feature extraction is the important breakthrough to solve this problem. Around the target of scene classification, this thesis uses scene classification algorithm based on local feature extraction as main research content. Then we research local feature extraction algorithm and feature coding algorithm. Research work of this thesis mainly focuses on the following four aspects:1. After studying dense SIFT feature extraction algorithm, then we improve it to solve the problem existing in feature vector normalization. We caculate several gradient direction responses, then find the main direction and normalize the descriptors. So the resulting feature descriptors have a certain rotation invariance, and the robustness of feature extraction algorithm has been improved.2. After studying covariance matrix feature extraction algorithm, then combined with dense grid sampling frame, we improve it to solve the problem existing in huge computation. Griding the imags, then calculating covariance matrix descriptors respectively. The improved algorithm reduces the calculation amount and ensure the accuracy of the extracted features at the same time.3. After studying local-constraint linear coding, we improve it to solve the problem existing in ambiguity of codebook mapping. we add criterion threshold, which make the coding algorithm has certain adaptability, during the processing of coding. The improved algorithm make codebook reconstruct feature vector more accurate. And it improves classification performance and reduces the coding time at the same time.4. Using the improved feature extraction algorithm and feature coding algorithm, we design two kinds of scene classification algorithm which are scene classification based on improved local featue linear coding and scene classification based on key point assisted linear coding. And they both improve classification performance. Experiments are going by using MATLAB, and the origin data is from 15-scenes dataset and 8-scenes dataset.