一种改进词袋模型的图像分类算法

A novel image classification algorithm based on improved BoW model

传统词袋模型易受视角、尺度和背景等因素干扰。本文对传统词袋模型进行了改进,提出一种基于角点检测和图论的感兴趣区域提取方法,以及结合高斯模糊隶属度函数选取视觉单词。首先,对图像进行角点检测,利用图论的方法划定ROI区域,然后对得到的ROI区域进行SIFT特征的提取并生成视觉词典,从而减少背景信息的影响。其次,引入高斯模糊隶属度函数改进图像视觉直方图的表示。在Caltech 100数据库上的实验结果表明,本文提出的方法相较于传统词袋模型,分类准确度提升了3%。

The performance of the traditional Bag of Words (BoW) model is vulnerable to viewpoint,scale and background.This paper improves the traditional BoW model,and proposes the region of interest (ROI) extraction method based on the corner detection and graph theory.Moreover,the paper selects the visual words via Gaussian fuzzy membership function.First,use corner detection for images,and determine the ROI with graph theory;then,extract the SIFT features in the ROI and generate the visual dictionary to decrease the influence of the background information;finally,introduce the Gaussian fuzzy member function to improve the representation of visual histogram of image.The experimental results on Caltech 100 database show that,the accuracy of the proposed method is 3% higher than the traditional BoW model.