Scene Classification Method Based On Statistical Feature Of Regional Research

With the rapid development of information technology, the image data is growing more and more, image retrieval based on image content is becoming increasingly important. Computer vision, as a kind of a new and hot research field. The main task of Computer Vision is to perceive the biological information by using the computer and related equipments. Making use of the computer to replace human brain data processing and displace. Scene classification is a key technology in computer vision. Realize artificial intelligence by inputting2D matrix and get high-level semantic interpretation. In this paper, we select the existing classic algorithm to analysis, research and implementation. To solve the problems found during the algorithm analysis. By combining the knowledge of digital image processing and pattern reorganization, We put forward a new algorithm based on super-pixels. The results of these algorithms have been improved more.The major research work is as follows:1. Elaborated the principle of bag of words (BOW), how to extract the characteristics of bag of words, and visual word extracting, the formation of the dictionary. Using six scenarios to test the classification accuracy of the BOW method and obtained classification accuracy the confusion matrix.2. PLSA (Probabilistic Latent Semantic Analysis) model-based scene understanding algorithms, a detailed analysis of the principle of the PLSA model, PLSA model to establish a text, word, semantic probability model. Using the EM algorithm to estimate the model parameters. Six types of scenes are used to test the PLSA model(judging function classifier), obtain the confusion matrix of classification accuracy.3. Presents a scene understanding (Super-pixels) classification technique based on super-pixel count, the use of super-pixel feature extraction algorithm to get the super-pixel blocks, each super-pixel block as a single unit, obtained by Gabor filtering super the texture features of the pixel block, and then calculate the color characteristics of the super-pixel blocks (first-order color moment, the eigenvectors of the super-pixel block) and the relative position information through the training of clustering new visual word form dictionary, and then use the trial of six scene classification and get the classification accuracy of the confusion matrix.There is an important difference among the three of these methods, the features extraction of the BOW and PLSA model are classic sparse regional characteristics (SIFT), and however, scene understanding based on the super-pixels algorithm is used in the regional statistical properties.