一种自动抽取图像中可判别区域的新方法

图像分割是图像处理中的一个难题，为了自动抽取图像中的可差别区域，提出了一种基于自组织图归约算法的区域抽取新方法，首先，利用包括颜色、纹理以及位置在内的多模特征抽算法，原始图像被转换成特征，接着，通过自组织映射学习算法，特征图映射成自组织图，然后，对自组织图实施归纳算法得到一族约简的自组织图谱系；最后，利用一个 综合的聚类有效性分析指标从约简的自组织图谱系中得到一个最优约简的自组织图，以此实现图像区域的分割，新方法的有效性通过两个评价实验得到了验证。

A New Approach to Automatic Extraction of Discriminant Regions in Image

Image segmentation is a well-known hard problem in image processing. In order to automatically extract discriminant regions from an image, this paper presents a novel method of region extraction, which is based on a SOM (self-organizing map) reduction algorithm presented in the paper. Firstly, according to a multi-feature extraction algorithm, the raw image is transformed into a feature map, in which each feature vector consists of three sub-features: 1) color feature-dominant color of a sub-region, 2) texture feature-MRSAR parameters of a sub-region, 3) and position feature-center coordinate of a sub-region. Secondly, SOM training algorithm is performed against the feature map generated at the first step. A self-organizing map, in which the number of units is much smaller than that of feature vectors in the feature map, is created after SOM training. SOM training establishes a relationship between units in the SOM and feature vectors in the feature map. Those feature vectors, which are close with each other at the feature space, may map to the same unit of the SOM. Then, a family of reduced self-organizing maps is produced using a two-phase reduction algorithm of SOM. At the first phase, the unit, which has the least number of feature vectors at the map, will be reduced. At the second phase, two units, which are nearest at the feature space, will be merged. Those feature vectors mapping to the reduced unit will re-map to other neighbouring units according to a BMU match rule. Finally, in order to select an optimum one from a series of reduced self-organizing maps, an unsupervised cluster-validity analysis is performed. Pixels in the raw image can be grouped into different discriminant regions according to the relationship between the relevant feature map and the optimum reduced self-organizing map. At last, two evaluation experiments are given to verify the effectiveness of the new method.