Image segmentation based on self-organizing dynamic neural network

Image segmentation is critical to image processing and pattern recognition, while feature space clustering is an important method for unsupervised image segmentation. The method assumes that image feature is a constant property of the surface of each object to be segmented within the image and the image feature could be mapped into a certain geometrical space called feature space. Meanwhile, the method also assumes that different objects present in the image will manifest themselves as different clusters in the feature space. Therefore, the problem of segmenting the objects of an image can be viewed as that of finding the mapping clusters in the feature space. Using a proposed novel competitive-learning-based neural network clustering algorithm to cluster image feature space, an unsupervised image segmentation method is realized. The proposed clustering algorithm, self-organizing dynamic neural net (SODNN), is possible to dynamically grow swiftly and adjust the size of neural net more accurately, so it is able to find the number of clusters according to the input data pattern and get more stable and accurate result. Here, the SODNN algorithm is utilized and the color and position features are combined to carry out the image segmentation. The presented results show better consistency between the segmentation by proposed methods and the segmentation by human.