基于在线被动-主动学习的多特征融合图像分类

Classification for Multi-Feature Fusion Images Based on Online Passive-Aggressive Algorithm

提出了基于在线被动-主动学习的多视觉特征自主加权组合算法。该算法在模型训练阶段预先依据视觉特征与图像类别之间的相互关系赋予恰当的权值,减少了多特征组合的计算复杂度。通过推导出在线被动-主动学习算法的闭式解,提出的算法在满足确保图像分类准确度的同时,提高了多特征组合的执行效率,降低了基于直方图交核学习算法的计算复杂度。与多核学习算法相比,基于在线被动-主动学习的多特征融合图像分类算法在保持图像分类准确度的情况下,所需的计算时间只有多核学习算法的10%左右。

An algorithm to learn the weights of combining multiple features for image classification adaptively based on a modified online passive-aggressive （OPA） algorithm was proposed. A closed-form solution for the modified OPA algorithm under the histogram intersection kernel was derived. The proposed algorithm can improve the classification accuracy and reduce the computational complexity simultaneously. Compared with muhiple kernels learning （MKL） methods, the proposed algorithm maintains similar classification accuracy to MKL, and the modified OPA algorithm takes about 10% running time of MKL. Experiments demonstrate that the proposed algorithm： 1 ） maintains equal classification accuracy to MKL; 2） reduces the computational complexity of the histogram intersection kernel-based classifiers significantly.