Image filtering with multilayered cellularly connected evolutionary neural networks
Junji Otsuka, Tomoharu Nagao · 2012
This paper presents a study of automatic construction of image filters with a novel multilayered cellular network. Each layer of the proposed network is Cellular Real Valued Flexibly Connected Neural Network (CRFCN), the neural network model for automatic construction of image filters we previously proposed. CRFCN consists of a regular array of evolutionary neural networks called Real Valued Flexibly Connected Neural Network (RFCN). In the previous work, single-layer CRFCN showed its good performance of image filtering. However in complex image processing, it is often effective to decompose the processing into a sequence of partial processing to achieve the target task. Hence, to enable the model to split up a complex task into partial tasks, we propose multilayered CRFCN: series-connected single-layer CRFCNs. The key ideas of multilayered CRFCN are (1) setting each layer to the original input image to prevent from lacking important information through multilayered processing, and (2) construction of the layers one by one efficiently. We apply multilayered CRFCN to three different image filtering of region extraction in comparison with comparative methods, and show its performance.