Neural network approach to unsupervised image segmentation

This paper proposes a new neural network model UMAN to perform unsupervised image segmentation. In the neural network, the generalized information entropy is used as the quantitative description and measurement of the system stability and asymptotication, and the disadvantage of generalized energy function is avoided. The improved Kohonen non-linear mapping structure not only contrasts the clustering features, but also reduces the redundant information. In the network, the internal layer and node numbers are determined dynamically by the system. The interaction and a prior knowledge are not required. The unsupervised self-learning function expresses the characteristics of the low-level visual information processing. The UMAN model can process various types of images and has strong adaptability. Experimental results show that the model and its algorithm are efficient, practical and robust.