Ping Zhong; Runsheng Wang;
Pattern Recognition, 2006. ICPR 2006. 18th International Conference on
Volume: 3
Digital Object Identifier: 10.1109/ICPR.2006.876
Publication Year: 2006 , Page(s): 160 - 163
Cited by: 1

百优得主钟平老师的处女作：结合CRF与MRF做城市区域检测。

一、流程图

二、特征提取

三类特征：线长统计（见上一博文）、梯度幅值、梯度方向

1、梯度幅值

采用梯度幅值的均值、p阶矩、熵、、能量作intrascale特征，作尺度间特征

2、梯度方向

依据：If the window Wc contains a smooth patch, the gradients will be very small and the mean of the histogram over all the bins will also be small. On the other hand, if Wc contains a textured region, the histogram will have approximately uniformly distributed bin magnitudes. Finally, if Wc contains many straight lines embedded in the urban area, a few bins will have significant peaks in the histogram.如果窗口是光滑区域，则梯度的均值和方差都很小，若为纹理区域则呈近均匀分布，而城市区域则会出现显著的峰值。

三、参数估计与推断

首先采用CRF对精确目标建模并获得初始检测结果,然后估计参数，最后进行推断。这一部分看不懂，先记下能看懂的，以后再补充理解。

1、参数估计

CRF参数即为特征向量各分量的权值，采用最大似然法进行估计；

MRF参数采用EM算法进行估计

2、推断

在CRF中采用Belief Propagation（BP）算法产生近似MAP标记；

在MRF中采用iterated Conditional Modes(ICM)方法在对象边界近似估计MAP