Weighted Natural Image Classification Based On SVM-KNN Adaptive Features

The problem of digital image classification plays a significant role in image processing research, and current research on image classification are mostly concentrated in the field of professional image, such as remote sensing and medical images, etc. These images often have low resolution, most of them are multi-grayscale image, texture is relatively simple and only a small amount of data. However, at this stage, for the study of complex natural images which human beings contact mostly and use widely in the daily life are relatively few. Therefore, this article will focus on natural images, and the content is natural images classification based on adaptive features-weighted SVM-KNN algorithm. For this research work, this article is mainly on the following three aspects:(1) This paper studies some classification algorithm in currently pattern recognition system, including:Bayesian classification algorithm, K nearest neighbor classification algorithm, support vector machines, artificial neural networks. The basic idea of each classification algorithms, classification steps and scope of applications are being introduced. The paper also have summarized the advantages and disadvantages of each kinds of classification algorithms, and in order to increase the performance advantages complementary and improve the classification accuracy, combined with the specific needs of the experiment, experiment choose combined classifier SVM-KNN;(2) Studied underlying visual feature extraction method which commonly used in pattern recognition system, including some color feature description algorithm for example color histogram, color space aggregation, color moment vector description; some texture feature description algorithm for example Tamura texture feature, gray symbiotic matrix and Gabor texture feature; some shape feature description algorithm for example Fourier shape descriptor, boundary characteristics, geometric parameters, etc. Each algorithm steps of the description method was introduced in detail, and each method has carried on the detailed classification experiment. Selected72d color histogram feature and40d Gabor texture features as the low-level visual features of the experiment. Summary and compiled several common methods of high-dimensional data visualization, represented the color and texture feature in parallel coordinates, found strong distinction between each different class;(3) Proposed a new classification method for natural images based on feature-weighted SVM-KNN algorithm. Combined this two classifiers, so that it not only can improve the accuracy but also save the operation cost. At the same time, new algorithm improved the KNN on two hands. On the one hand, considering color and texture feature make different contribution on natural images classification, so characterized the features by giving them different weights. Combining genetic algorithm and gradient descent of the iterative algorithm to get this optimal weight, and in the algorithm user can change the classification accuracy according to their own actual need.On the other hand, use included angle cosine instead of Euclidean distance in the original algorithm to measure the similarity between the sample and the sample under test. Experimental data show that compared with the original single classification method, the new algorithm which proposed in the paper is increased by2to7%on the average classification accuracy.