论文部分内容阅读
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。
The task of accurately detecting many kinds of targets from high-resolution remote sensing image data sets higher requirements for the detection speed and model training time. In this paper, a MKL_mRVM method is proposed. This method uses the fast edge-likelihood maximization algorithm to directly compute the mRVM classifier decision-making function, which avoids the traditional RVM iteratively computing the Hessian matrix of the objective function, and because there is no need to construct a series of two categories Which can shorten the training time of many kinds of models. At the same time, a number of basic kernels are introduced into the multi-class model, the cross-validation method is used to determine the basic weight of the kernel in the training process, and the accuracy of the classifier is checked on the random- Making the classification model with the highest accuracy as the weight of the optimization results. Experimental results show that this method can shorten the model training time effectively while keeping the sparsity of solution.