基于ReliefF特征加权和KNN的自然图像分类方法

Classification Method of Feature Weighted for Natural Images Based on Relief F and K-nearest Neighbors

为了对自然图像有效准确地分类,提出了一种对图像低层特征和KNN分类算法中的近邻样本分别进行加权的分类方法。针对不同类别图像的视觉特征的差异,通过Relief F算法计算训练集中每个类别的特征权值,利用此权值来改进待测图像与训练集中图像的距离度量;按照不同近邻到待测样本的距离远近,为不同近邻赋予权值来改进KNN算法在类别决策上的不足。实验结果表明该方法较传统KNN和特征加权KNN方法,准确性提高且对不同K值具有良好的鲁棒性。

In order to classify the natural images more effectively and accurately, a classification method weigh images feature and the nearest neighbors of KNN is proposed. Since diverse categories images have different visual features, ReliefF is used to obtain the feature weight vector of each category in training set for weighing the distance between test images and training images ; different weights are given for the K-nearest neighbors according to the distance to training images, so that the weakness of traditional KNN at the classification decisions is overcome effectively. Compared with the traditional KNN and feature-weighted KNN, the experimental result shows that this method has more accuracy and strong robustness for the number of the nearest neighbors.