Research On Application Of Cellular Neural Networks In Video Motion Objects Segmentation

The reason of research on application of cellular neural networks (CNN) in video motion segmentation is the high speed concurrent operation character of CNN and the network construction which is quite fit for image processing. Up to now, the whole operational speed of existent CNN chips has reached Tera magnitude, which can great improve the speed and real time of video processing.In this paper, the research motive of cellular neural networks application in video motion segmentation is discussed. After reviewed and compared the existing video segmentation methods, the previous application of CNN in that field is introduced. According to the network construction character and mathematical physics model of CNN, the CNN dynamic range and stability is analyzed in detail, and the quantization method of CNN input and output in image processing and the basic principle of CNN applied to image processing is introduced. Aim at video sequences with static background, the difference merged image algorithm based on CNN is presented. In order to realize the algorithm, five CNN templates are constructed. Aim at the estimation of motion field in video sequences, the more perfect optical flow field algorithm based on CNN is presented, and the CNN templates of difference algorithm on space and time domain of video sequences are constructed, at the same time the 8-bit quantization method of CNN used for motion field analysis is introduced. Aim at video sequences with dynamic background, the CNN video motion segmentation algorithm based on optical flow field threshold is presented.The main novel contributions of this paper are as follow:1. Difference merged image algorithm based on CNN is presented. This algorithm is aim at motion segmentation of video sequences with static background, and it belongs to one of history area detection methods. Compared with some existing similar algorithms, because this algorithm orient gray-scale images directly, it can obtain more motion information and improve the accuracy and semantic character of motion objects segmented from video sequences.2. Five CNN templates are constructed, which can be used for image and video processing. In order to realize the difference merged image algorithm based on CNN, five CNN templates are constructed, which are negative template, addition template, edge detection template, dilated template and composition template. Those templates not only can be used for this algorithm, but also are fit for other...