形状和颜色混合不变量在图像检索中的应用

多视觉特征的图像检索是当前基于内容的图像检索领域的重要方向.已有的多特征的检索主要通过线性加权的方法对特征进行组合,但这种组合方式仅实现了代数意义上的合并,未能真正利用和发掘特征间存在的相互关系,并且权重值不容易确定,检索结果易受权重值的影响.针对这一问题,提出一种形状-颜色混合不变特征的构造方法,特征提取的过程包含同时对形状、颜色信息的抽取,直接构造出能够同时对形状仿射变换和颜色对角-偏移变换具有不变性的特征,也称作形状-颜色矩不变量.首先分别在图像的2维几何空间、3维颜色空间定义形状核、颜色核,然后对形状核、颜色核的乘积进行多重积分,最后做规范化,就得到一个不变量.理论上,通过选择不同的形状核、颜色核可以推导出无穷多的不变量.实验结果表明,该方法优于加权组合特征的方法;与加权特征、局部特征相比,形状-颜色矩不变量对于同一物体不同成像条件下的近复制图像、整体属性相似的图像、大体类似的物体图像等表现出较高的检索性能及效率.

Application of combined shape and color invariants to image retrieval

Image retrieval based on multiple visual features is one of most important direction in the field of content-based image retrieval. Existing retrieval algorithms are based on multiple features mainly combining features through linear weighting. However, this just achieves combinations in the sense of algebra, and the retrieval result is susceptible to the choice of the weights. To solve this problem, including the extraction of shape and color information in the feature extraction procedure, we propose a method to construct combined shape-color features, called shape-color moment invariants, which are invariant to both shape affine transformation and color diagonal-offset transformations. First, we define the shape core and color core in the two-dimensionol geometric space and three-dimensionol color space respectively. Then multiple integrals to the product of shape core and color core are produced. Finally, we get an invariant feature by normalization. Theoretically, this method can derive an infinite number of invariant features while choosing different shape cores or color cores. The experimental results show that this method is distinctive to existing linear weighting method, and has better performance and efficiency when retrieving near-duplicated images,images with similar property,broadly similar images etc.