Research On Cellular Neural Networks And Its Application On Edge Detection

Cellular neural network (CNN) was first proposed by Chua and Yang in1988. Thisstructure combines the advantages of artificial neural network (ANN) and cellularautomata (CA). CNN is a variant of ANN，which is local connection of each other. Thischaracter facilitates high speed parallel processing ability. The whole network iscomposed of large-scale nonlinear analog circuit and can easily implement on VISI. Sofar, CNN has received great success in its application on image processing, and its mostchallenge work is finding a series of proper templates to fit the given task. There aretwo popular methods to find the templates. One is cloning templates. To get cloningtemplates, it is always using intelligence algorithm, such as PSO or solving linearmatrix inequality (LMI). These two methods are all rely on a samples space. Thedifference of them is that intelligence algorithm to training samples in the samples space.And LMI method is abstracting a series of linear matrix from the given samples, then,solving them. The other is adaptive templates. It is always according to the specialty ofthe given tasks to directly design the templates.In this paper, CNN is applied to color image processing. In the templates designprocess, an adaptive templates designing algorithm is proposed. It takes the specialty ofedge detection to simplify the designing process. In general, the main contributions ofthis paper can be described as follows:①So far the most popular method in CNN to image processing is to find cloningtemplate. So, this paper, apply PSO and LMI to find CNN cloning templates in edgedetection.②There is few method addresses color edge detection using CNN. It is hard to finda CNN template model to specify color space. And there neither be an accepted methodto find cloning template in color space. In this research, application CNN to color spaceis first proposed. In edge detection, the change of pixels receives more attention than thevalues of pixels themselves. Take advantage of this character, the multiple CNNstructure is avoided in color edge detection.③It employs the achievement of human vision system to design adaptive threshold.Weber-Fechner law points out that the minimal color changes that can be detected byhuman vision system vary with the change of the background. And the reference paper[40] defined a specific function, which indicates the mathematical relationship of minimal detected change and background. Combining the above achievement anadaptive threshold is got.④Some modifications are to be done on CNN structure. This modification isinspired by the observation on the method in [45，46]. In [45，46], shadow detectionalgorithm has been proposed. It assumes feedback part to be a fix const. In this paper, itemploys the threshold that discussed above to be decision condition. This decisioncondition decides the value of the feedback template. These templates carry the changeinformation of the inputting image.Experimental results show the efficiency of the proposed algorithm. And it hashigh speed of convergence.