基于稀疏重构残差和随机森林的集成分类算法

Ensemble classification method based on sparse reconstruction residuals and random forest

传统的基于稀疏表示的图像分类算法,通常根据稀疏重构后类残差向量的l2范数得到分类判决.在复杂情况下,各类残差向量l2的范数差别可能并不明显,从而导致分类器作出错误判决.提出了一种基于稀疏表示和随机森林的集成分类方法,通过稀疏表达字典对图像进行重构,提取各类残差图像的l2范数组成特征向量,并引入随机森林进行分类判决,有效地提升了算法基于类残差向量的判决能力.在手写数字数据库MNIST上的实验结果表明,在训练样本数较少的情况下,提出的基于稀疏表示和随机森林的集成学习分类方法与目前主流的SVM分类方法及随机森林方法进行比较,识别率有较为明显的提高,具有良好的鲁棒性.

Based on the sparse representation computed by l2-minimization and ensemble learning,we propose a general classification algorithm for image classification.This new framework provides new insights into two crucial issues in image classification：feature extraction and classification accuracy.Since it was proposed,random forest has become a well-known data analysis method,and it has been applied to a wide variety of scientific areas.As the random forest classification has a good performance and high stability on classification,in this paper,we choose random forest as an ensemble learning classifier.The classifier based on sparse representation classified the test sample by calculate its l2 norm of residual vector between its real values and its reconstructed values.While in some cases,due to the difference of the residuals are very small,it is hard to decide the right class that the test sample belongs.We have proposed a reconstruction algorithm of sparse representation to extract image features and classify the images by random forest classifier.First,a learning dictionary is obtained based on the trained image data set.We generate a sparse vector on the over-complete dictionary,and then calculate the residuals between the real values and the reconstructed values of the training samples.The residual vector is used as the training sample of the random＆amp;nbsp;forest classifier.Finally the image is classified by the trained random forest classifier.Random forests are respectively constructed based on residuals,and the classification result is decided by voting strategy.Our Experiments use the standard digital database MNIST as the image recognition database.The recognition rate of the method proposed in this paper is obviously prior to some other popular classification methods,such as SVM.We use MATLAB to finish the research experiment.The experimental results indicate that the method we proposed has better performance than methods based on random forest and sparse representation respectively.Besides,this method has the stability of the result of the classification and good noise robustness.