基于四叉树和交叉熵的面向对象图像分割方法

Object-oriented image segmentation method based on cross-entropy and quad-tree

为了解决光照不均的非均匀图像分割问题,提出了基于面向对象思想的图像分割算法.针对非均匀图像的特点,以四叉树作为分析结构,在Shannon熵上推导出子集熵与全集熵的关系,作为图像的面向对象描述.基于此关系,充分考虑非均匀图像子集的局部灰度分布,最小化子集与全集的交叉熵,抹去子集的局部灰度偏移特征,从而得到分割阈值与局部灰度分布相关的分割方法.实验表明,相比常用的动态阈值算法,该算法具有运算量少、分割结果自适应性好的特点.

For dealing with some inhomogeneous images, due to the factors of non-homogenous illumination, a novel segmentation based on object-oriented theory was provided. According to the features of the inhomogeneous images, the relationship between entropy of entire set and that of subsets was deduced based on the structure of quad-tree, which was analyzed as the object-oriented description. The shift distribution was removed through minimizing the cross-entropy between the subsets and entire set from the relationship, and a new entire set was reconstructed according to the removed subsets, which includes more particular information of the original image. The proposed thresholds were dependent on the local grayscale distribution of the subsets. Experiments show that the algorithm has the advantages of less computation and better segmentation than other usual algorithms of adaptive threshold surface.