Feature Extraction Algorithm Optimization And Implementation Used For The Scene Image Classification

With the development of network technology, image is used more and more in people’s life. How quickly and accurately to classify a large number of image and to provide more accurate information for the more complex image processing become a research focus. Based on application of Bag of Words (BoW) model in image classification, we study the "human"&"inhuman" scene image classification and optimize image feature extraction.The problem of the traditional scale invariant feature transform (SIFT) in the extraction of local features is limited key point and high complexity. In order to solve this problem, we use Dense-SIFT algorithm to optimize the local feature extraction. By using a rectangular window replace Gaussian window on feature point extraction and generation of descriptors, can get more key points and reduce calculation complexity.Based on studying the high-level image feature generation, In this paper we first select the optimal parameters for the size of the visual dictionary, ensure the accuracy of image classification. At the same time, we deeply studied principle of the Spatial Pyramid Matching (SPM) model, combined with Average-pooling and Max-pooling algorithm. Through the experimental comparison of two pooling algorithms, we found that max-pooling have stronger robustness for feature changes in "human"&"inhuman" scene image classification.Because of the local feature can’t be fully represented image global features, we combine global color feature with local feature as image final feature, and use color feature in the SPM model as well. Optimal level of color feature in SPM model was determined by experiments in this paper. Comparing before adding global color feature classification results, the optimized algorithm can effectively improve the performance of color image classification. We use the complete optimized model and compare the classification results with before, the optimized model have higher classification accuracy rate, so we lay a foundation for human image segmentation and image retrieval.