多核多类关联向量机的高分辨率影像目标检测  

Multiclass objects detection in high-resolution remote sensing images using MKL_mRVM

作　　者：李湘眷[1] 孙皓[2,3] 王洪伟[4] 王彩玲[1] 

机构地区：[1]西安石油大学,西安710065 [2]中国科学院空间信息处理与应用系统技术重点实验室,北京100190 [3]中国科学院电子学研究所,北京100190 [4]中国人民武装警察部队工程大学,西安710078

出　　处：《测绘科学》2014年第12期128-133,137,共7页Science of Surveying and Mapping

基　　金：国家自然科学基金(41301480;41301382);西安石油大学青年科技创新基金项目(2013BS014)

摘　　要：从高分辨率遥感图像数据中准确检测多类目标的任务对于检测速度和模型训练时间提出了较高的要求。文章提出了一种MKL_mRVM方法:该方法采用基于快速边缘似然最大算法直接计算mRVM分类器的决策函数,避免了传统RVM重复计算目标函数Hessian矩阵的过程,并且因为不需要构造一系列两类分类器,缩短了多类模型的训练时间;同时,将多个基础核引入多类模型,训练过程中采用交叉验证方法确定基础核权重,在随机分出的确认集上检验分类器的精度,选取使得分类模型精度最高的值作为权重的优化结果。实验结果表明,该方法能够在保持解的稀疏性的前提下,有效地缩短模型训练时间。Accurate objects detection from massive data in the high-resolution remote sensing images put forward higher demands for inspection speed and model training time. A MKL _ mRVM algorithm was proposed in the paper. Fast marginal likelihood maximization was utilized to directly train the decision function of multi-class classifier, avoiding the necessity of repeatedly computing Hessian matrix in traditional RVM. Furthermore, the proposed method reduced training time greatly without constructing lots of twoclass classifiers. A cross validation strategy was applied to optimize the kernel weights. Then the accuracy on validation sets randomly selected from training sets was tested, and weights corresponding to the maximum accuracy were chosen as the final result. Experiments results showed that the proposed method could not only keep the sparsity of the solution but also reduce the model training time.

关 键 词：高分辨率遥感图像 多类关联向量机 多核学习 多类目标检测 

分 类 号：TP75[自动化与计算机技术—检测技术与自动化装置]