Abstract: Using single color or texture feature can not achieve satisfied image retrieval effect. In order to solve the problem, a new image retrieval method which combined the color and texture features was proposed. The color component was retrieved from both micro and macro parts. In the micro part, the color histogram was used to describe the proportion of pixels in each color of the whole image. In the macro part, color entropy and bit-plane entropy were used to process the image respectively in order to exclude the obvious different images from the aimed picture. The first 4 layers with distinct features were selected from the bit-plane entropy part, and different weight value was given for bit-plane entropy of each layer. Finally, according to the each pixel color value and the angle value of the defined five basic texture elements, combined with color feature, image retrieval was achieved. The experimental results on Corel-1000 dataset show that, compared with unweighted bit-plane entropy, the average precision and recall of weighted bit-plane entropy were increased by 10.01 percentage points and 1.2 percentage points respectively. Moreover, compared with Structure Elements' Descriptor (SED) method for texture feature only, the proposed method combining the color and texture feature improved the average precision and recall by 4.3 percentage points and 2.1 percentage points respectively on Corel-10000 dataset. The proposed method can effectively improve the effectiveness of image retrieval.

Key words: color histogram, color entropy, bit-plane entropy, texture unit, image retrieval

摘要： 针对仅使用单一颜色或纹理特征并不能达到较好的图像检索效果的问题,提出了一种结合颜色和纹理特征的图像检索算法。首先,颜色微观部分利用颜色直方图,刻画每种颜色的像素占整个图像的比例;然后,宏观部分应用颜色熵和位平面熵分别对图像处理,其中位平面熵取特征较明显的前4层,并对每层的位平面熵加权;最后,根据定义的五种基本纹理结构基元中各像素点的颜色值和角度值,结合颜色特征,实现图像检索。实验结果表明,加权位平面熵和不加权位平面熵比较,在Corel-1000数据集上平均查准率和平均查全率分别提高10.01个百分点和1.2个百分点。结合颜色和纹理特征的图像检索算法与仅表现纹理特征的结构元素描述(SED)方法相比,在Corel-10000数据集上平均查准率和平均查全率分别提高4.3个百分点和2.1个百分点,有效地提高了图像检索效果。

关键词: 颜色直方图, 颜色熵, 位平面熵, 纹理单元, 图像检索