基于自组织特征映射的图像分割算法研究

Application of Kohonen networks for image segmentation

提出了一种基于自组织特征映射的图像分割算法,实现了计算机对图像的初步理解,从而在某种程度上模拟了生物的初级视觉功能.通过分析研究Kohonen网络的自组织特征映射过程,构造了基于Kohonen网络的图像分割神经网络方法,应用自组织特征映射方法将原始图像分割为有序化的相关特征区域.最后结合图像分割的特点对算法进行了改进,结合有监督的学习算法,使得图像的分割最终在先验知识的指导下进行.实验结果表明将Kohonen网络应用于图像分割使得算法具有很强的自适应性,能够在很大程度上避免背景及噪声对分割的影响.

A problem of image segmentation, which is due to problems of classification, was done in a supervised manner in general. A new image segmentation algorithm based on the self-organizing feature map （SOFM） theory was proposed, since Self-organizing maps can learn to recognize groups of similar input vectors in such a way that neurons being physically close by in the neuron layer are able to respond to similar input vectors. Thus row images can be classified by computer automatically, which is to recognize image and analog the primary visual process. A row image is separated into a serial of correlated regions ordered by their regional feathers through SOFM method. In order to extract the object region out of the raw image, the learning vector quantification （LVQ） algorithm is introduced into the training process, and the experiment result showed that the neural network classifier proposed is adaptive and robust to the background noise in image segmentation.