融合空间信息LDA的视觉对象识别研究

LDA Model Combined Spatial Information for Visual Object Recognition Research

近年来国内外有很多学者将广泛用于自然语言处理的LDA模型引入视觉对象识别,对象分割,场景分类等应用中。LDA模型是产生式模型,所以必然存在产生式模型共有的弊端,即假设每个视觉词汇所对应主题的产生是条件独立的。根据图像本身的特征,图像的空间信息对图像物体识别起了很大的作用,一个视觉词汇主题的生成过程是受其相邻视觉词汇主题所影响的,为了提高图像视觉词汇的主题分配正确率,提出融合空间信息LDA模型,即融合条件随机场的LDA模型,从而在图像的局部主题标签上融合二维图像空间信息,既避免空间信息的丢失,同时可以提高其视觉词汇的主题分配正确率。论文主要研究内容:首先对LDA模型进行改进,并在该模型中引入条件随机场,同时推导出使用期望最大化算法确定的模型参数。该论文提出使用条件随机场获得图像的二维空间信息,将产生式模型和判别式模型融合,在增强由图像本身性质所决定的相邻区域主体标签的空间关联性的同时,也提高了视觉对象识别的精确度,完成图像的自动标注。

In recently years, many scholars introduce the LDA model which is widely used in nature language processing into visual object recognition, object segmentation, scene classifieat/on and so on. LDA model is a novel generative model, so there must be common defect between generative models that it assumes latent topic assignments of different visual words are conditionally independent, According to the eharacteristics Of images, spatial information of the images plays an important role in image object recognition, that is to say, the generation process of the latent topics given the visual words is influenced by its adjacent visual words＇ latent topics. In order to improve the accuracy of the distribution of the topics given the visual words, the paper proposes the LDA model combined spatial information, namely LDA model combined CRF, which is fused the 2D image spatial information in the local latent topic label to avoid the Ioss of spatial information and can improve the ,accuracy of the distribution of the latent topics. The main research Contents of this paper： firstly, improve the LDA model, and combine the conditional random field into LDA model, and derive the model parameters using the corresponding EM algorithm. This paper uses the conditional random fields for getting 2D spatial information of the images ; combines the generative model and the discriminative model. The paper enhances the spatial correlation of the latent topic labels of the adjacent visual words determined by the images＇ nature characteristic, at the same time, improves the recognition rate of the visual objects.