A Fast Image Classification Method Based On Matching Similarity And TF-IDF Value Of Region

With the rapid development of computer network and multimedia technology, image, as an important carrier of information, is increasingly being used to record and transmit information. How to implement information classification by analyzing the characteristics of image has attracted widespread attentions. Image contains a lot of low-level visual features information like color, texture, shape, spatial relationship and so on, which are the intuitive feelings of people to the image content information; High-level content semantic characteristics contained in the image belongs to the subjective understanding of image content information. However, the low-level visual features and high-level semantic content contained in the image have bigger difference, which is called "semantic gap". How to eliminate the "semantic gap" is the key to solve the problem of image classification.Traditional image classification methods are mainly based on the overall color related statistics and semantic contents contained in the image. The poor difference of the overall color related statistics and semantic contents contained in the image results in poorer actual effect in the process of image classification. Therefore, this paper proposes a fast image classification method based on matching similarity and TF-IDF value of region, and the main research work are as follows:(1) Divide all images in the standard image data set into some a number of fixed size image regions (including1*w pixels), and extract the RGB value of each pixel in each image region as the characteristic information. This method takes into account the image pixel distribution and relative location information, can effectively avoid wrong classification caused by the similar overall color information of images.(2)Extract the image feature information of all different regions in the standard image data set, calculate the matching similarity between each image region, build the standard image library to the image data set, and set the k-means clustering index for the regional characteristics in the standard image library. It improves the query speed of matching different regional characteristics, contributes to the realization of fast image classification and improves the efficiency of classification.(3) Introduce and analyze the classic TF-IDF algorithm in the process of text categorization, combine the proposed image regional feature extraction and image matching similarity calculation method, and propose the TF-IDF value calculation method of different image region corresponding to each image categories. In the process of the actual image classification, the different image region TF-IDF value to different category can be used to quantitatively characterize the importance to decide the image classification result.(4)Use the different image region matching similarity in the standard image regional characteristics library and different image region TF-IDF value to different category, and further update the information of the image features and different image region TF-IDF value to different category in the standard image library.(5) In the process of the actual image classification, to classify a new image, first implement the regional segmentation and feature extraction of image region, then retrieval and match the standard image feature library, construct the high-dimensional image feature vector which is used to describe the different regions of the image are divided into different categories, then implement the high dimensional feature vector dimension reduction and build the confidence characteristic vector of characterization of the image is divided into different categories, last take the feature vector of the highest degree of confidence as the final classification result of the image.