条件随机场模型的场景描述

Scene description based on the conditional random fields model

提出一种基于条件随机场模型的场景描述方法,条件随机场模型直接对描述目标的后验概率建模,不但能融合多类特征,还具有联系上下文信息的能力,这使得CRF模型在场景描述中能获得更准确的描述结果。将图像分成m×n大小的矩形块,通过多类特征提取,分别提取图像中每一矩形块的颜色特征、纹理特征、位置特征,通过K-means算法对特征进行聚类,并按照矩形块的位置组成特征向量,用CRF模型对特征向量建模,通过训练获取模型的参数估计,最终利用MPM算法进行模型推断,获取场景描述。实验结果表明本文方法能较准确地进行场景描述。

A method of scene description based on the conditional random field model is presented in this paper. The conditional random fields models the posterior directly, so that it can exploit several types of features, and has the ability to contact context information. Therefore, the CRF model in the scene description can get a more accurate description of the results. In this paper, the images are divided into rectangular blocks with a size of m ~ n. The color feature, texture fea ture, and location feature for each rectangular block are extracted through multiclass features extraction. These features are clustered by the Kmeans algorithm, and then the feature vector is composed of the features clustered by Kmeans in accordance with the position of the rectangle. The feature vector is modeled by the CRF model. The model parameters are estimated through training. We use the MPM algorithm for model inference to get the scene description. The experimental results show a higher accuraly of the method presented in this paper for scene description.