Image Classification Based On Complex Networks

Image classification is an image processing method based on the characteristics of the digital image, according to the difference between the characteristics of the image, to make the target image into different categories. With the development of computer technology and Internet, using artificial intelligence, machine learning and other methods to achieve image classification is a hot issue in recent years.In the research of image classification, feature extraction and feature coding is the study of a hot and difficult issue. Feature extraction is the basis of image classification, and the completeness of feature in feature extraction is the basis of correct classification. Feature coding is a statistical and analysis for image features, which directly affects the accuracy of image classification.In this paper, the image classification technology is studied from two aspects: feature extraction and feature coding.First, in the feature extraction process proposed a new feature extraction method based on complex network. In the image, we can use feature points as the network nodes, the relationship between the features points as the weight. Complex network structure and image feature extraction are prepared adequately for the image classification. Analysis and judgment the common feature extraction algorithm, this thesis selected the SIFT feature point as the network nodes, and then get the weight of by correlation coefficient. Secondly, used the statistical constants to descript the feature network characterize and make sure the expression of image information complete. Then, in feature coding process, in order to suppress the instability of local linear constrained coding proposed a new coding method, that is, through the increase in non-negative constraint conditions, in order to enhance the stability of the coding algorithm. In the stage of feature pooling, the maximum value is used to descript the expression of the image. Final, the classifier is used to classify the image.Experimental results show that the proposed algorithm can effectively classify image. On VOC Action classification dataset, classification accuracy can be up to 97.00%. Therefore, it is a more efficient image classification algorithm.