Research Of Image Segmentation On Cellular Neural Networks

A cellular neural network (CNN) is composed of a number of units (cells) with local interconnections. Each cell in a CNN is formed by linear and nonlinear circuit elements. Due to the fact that CNNs are well suited for VLSI implementation and parallel computing, they have been used to tackle image processing tasks and some complex tasks that traditional methods cannot do well (say, object segmentation using active contour). The application of CNN is determined by its templates, which determine the dynamics of the network. We lay primary emphasis on the template design and application of image segmentation based on CNN in this thesis. We design templates of CNN dedicated to image gradient, edge detecting and image segmentation of active contour. For the purpose of image segmentation, we realize the GVF field using discrete-time cellular neural network (DTCNN).The main work in this thesis includes:1. Review the basic concepts, background, state of the art and hardware implementation of CNNs; and describe the purpose, significance and common technology of image segmentation.2. Analyze in detail the dynamic range and stability of a CNN, indicate that the stability of a CNN is closely related to the central element of its feedback template, and explain the quantization method of CNN input and output in image processing and how a CNN can be applied to image processing.3. Discuss the study method and algebraic design method of a CNN template. Derive the learning rule based on the gradient descent method for template design of CNN image processing, give the corresponding training algorithm and simulation results, and confirm that the method is feasible. Under an algebra framework, study the design of CNN template for edge detecting, derive the value range of CNN model from common template under a series of conditions, and justify this method via simulation experiments and with values in the derived range.4. Propose an implementation method of gradient vector flow (GVF) field using multilayer CNN, which is combined with expanded and thinned model of CNN to conduct image segmentation of active contours. The proposed method is not only of low computational complexity, but also avoids the local minimum in CNN active contours. In image processing, the initial contour guided by GVF information of external image will be evolved until it reaches the desired position. Experimental results show that this method outperforms the one proposed by Vilarino, et al.