MCNNとSOMを用いた動画像の記銘と動的想起

抄録

Various association memory models have been proposed with artificial neural networks. For example, an interconnected network model proposed by Hopfield is able to recall stored patterns stably, a chaotic neural network (CNN) proposed by Adachi et al. is able to recall stored patterns dynamically. Kuremoto et al. proposed a multi-layer chaotic neural network (MCNN) with CNNs, which is able to recall multiple time series patterns orderly and dynamically. However, conventional association memory models used to be examined their association ability by experiments with simple binary patterns. In this paper, a novel association system is proposed to realize chaotic recollection of time series for video images using MCNN. In the proposed system, features of video images are extracted and clustered by Kohonen's self-organizing map (SOM), and those clustered feature maps are transformed to be binary images which are stored by MCNN with Hebbian Learning rule. In the recalling process, MCNN outputs time series patterns of different video images in the sense different features, and typical frame of images is able to be reproduced by the median feature vector. Dynamical and temporal association of the proposed system for the video images was confirmed by the experiment results.