一种基于不平衡样本集的摩托车识别算法

An algorithm based on imbalanced data sets for motorcycle recognition

提出了一种基于HSV(Hue-Satura-tion-Value)空间的Haar小波特征和多SVM(Support Vector Machine)分类器的摩托车识别算法,以解决因样本比例不平衡所导致的对摩托车识别性能差的问题.首先在HSV颜色空间基于无符号小波系数构造特征提取算法,然后对训练数据应用所提出的样本重构方法得到若干训练子集,基于各个训练子集训练相应的SVM分类器,识别时将各SVM的输出结果进行融合即可得到最终识别结果.实验结果表明:该方法识别性能高,鲁棒性好,对于受数据的不平衡性严重影响的对象识别具有较好的应用和推广价值.

A motorcycle recognition algorithm based on Haar wavelet features of HSV space and SVM ensembles is proposed to solve the poor performance of motorcycle recognition generated by imbalanced data. At first, a feature extraction algorithm based on unsigned wavelet coefficients of HSV space is presented. Then a reconstruction ap- proach is applied to training data so as to generate several subsets, and several SVM classifiers are trained based on all subsets respectively;the final recognition result is obtained by aggregating the outputs of all SVM classifiers. Experimental results demonstrate that the proposed recognition approach has better performance and robustness than the current methods and shows promising value and prospect for popularization especially on occasions where classi- fication performance is badly affected by imbalanced data in practical application.