中文摘要：

自顶向下的颜色注意力算法（CA）用patch的颜色注意力值当做patch形状特征的权重来进行图像表示,其中,如果当前patch的颜色属于本类内经常出现的颜色,此patch上所提取的形状特征就被赋予一个较大的权值,否则就赋予较小的权值.但是算法没有考虑到对象颜色的多样性.本文提出的方法认为,对象上的patch应该是那些在某类中经常出现并且在其他类别中很少出现的特征.为了提高CA算法的对象识别能力,提出了一种基于互信息的对象颜色选择方法,此方法选择与每类最相关的几种颜色作为对象颜色,通过优化目标函数用于最终确定有类区分的对象颜色个数.并且在对象表示的时候对估计到的对象上的patch赋予一致的高权值.实验采用Soccer,Flower 17和PASCAL VOC Challenge 2007三个图像集进行测评,实验结果表明该算法能够得到比较好的分类结果.

英文摘要：

Visual attention is effective in differentiating an object from its surroundings.Top-down color attention（CA）method is developed to use color to guide attention by means of a top down category-specific attention map.CA is proposed based on the assumption that the color often appears in a category is supposed to be the object color and the patches with these colors are assigned different large weights,and it assigns the weight of the patch shape feature based on the patch color for image representation.However,the diversity of object colors is not considered in this method,and the object patches with different colors are assigned different weights for image representation.Moreo-ver,the object patches cannot be distinguished by CA.We suppose that the object patches often appear in one category and seldom appear in the rest of the categories.To enhance the object recognition capability of CA,an object patch selection method is proposed based on mutual information between the category and the color word which measures their mutual dependence.An object is usually characterized by many colors in the real world;the most representative colors are the discriminative ones.We propose the discriminative color histogram which only preserves the most discriminative colors to evaluate the color difference between different two different categories.Ranking mutual information in descending order is useful to show the importance of each color for a certain category.The higher the mutual information between the category and a color word is,the more likely the color is to be the object color in this category.We rank the mutual information between colors and a category in descending order,and the colors corresponding to the top few highest pieces of mutual information are selected as the discriminative colors in this category.Our image representation is based BOW,object information in the histogram is important for classification and all the object patches are equally important for finding the object regions.To this end,we combin