Research And Application Of The New Algorithm For Image Classification And Recognition

With the rapid development of multimedia technology and Internet network, image classification has become hot research topic at home and abroad. Fast and high precision image classification algorithm has been the basic premise to apply to practice, so the research of image classification has great importance.Support vector domain description has advantages of strong expandability, low complexity and not high requirement in the training sample size; affinity propagation clustering is a deterministic classification algorithm, the classification of the multiple independent running results are generally very stable, and it has advantages of simpleness and efficiency, this two algorithms have been widely used in image classification.This text does research into the theory and knowledge of image classification, on this basis, two kinds of improved algorithm about image classification are proposed. First, remote sensing image classification algorithm which is based on improved support vector domain description is proposed; second, human face classification algorithm which is based on improved affinity propagation clustering is proposed. Main work as follows:(1) Improved algorithm of image classification which is based on support vector domain description is proposed mainly for remote sensing image classification. Firstly this text extracts spectral feature, wave texture feature and elevation feature of remote sensing image, and formed combination of feature vectors are proposed, which makes up for the inadequacy of the simple using spectral feature; then considering radial basis kernel function has shortcomings of large computation, weak generalization capability, this text introduces K-type kernel function and exponential radial basis kernel function on the basis of radial basis kernel function, which improves image classification precision; finally, considering different single kernel functions have different level influence on input feature vector, single kernel function in the expression differences of all kinds of feature vectors has certain defects, in order to weigh the various differences between feature vector, multi-kernel function is proposed on the basis of single kernel function. Results of simulation experiments show that improved single kernel and multi-kernel algorithm all improve the problem of finely spots and boundary pixels which are easy to be wrongly classified, and multi-kernel function improves classification precision more greatly than single kernel function.(2) Improved algorithm of image classification which is based on affinity propagation clustering is proposed mainly for human face classification. Firstly constriction factor is introduced to adjust the convergence coefficient and accelerate convergence of affinity propagation clustering; then validity index is embedded in iterative process of algorithm to supervise and guide algorithm to run in the direction for the best clustering quality. Conduct simulation experiment on human face classification and recognition with improved algorithm, experiment shows that improved algorithm can get more accurate clustering numbers than affinity propagation algorithm, Fowlkes Mallows value is improved, and the error rate reduces.