Multichannel Image Filter Based on FNN

摘要

A new filter for the multichannel image based on the technology of Fuzzy Neural Network (FNN) is proposed. Multichannel images obtained by remote imaging technology are often corrupted by noises such as gauss noise, impulse noise and speckle noise. Tradition linear filters and some nonlinear filters can't fulfill the task of removing the mixed noises clearly and quickly. The biology vision systems are sense to wide bind of light and their network structure can perform in parallel manner and filter mixed noises, which inspires to the design of FNN image filter. The neural network structure in the presented filter is good at learning from sample data and parallel calculating while the fuzzy mechanism embedded in the network can detect different patterns in order to remove noise and keep details and textures. The filter has three function layers. In the first layer, corrupted image is introduced into five fuzzy sets characterized by membership functions. In the middle layer, fuzzy reasoning hidden in the neural network detects the data pattern and indicated noise pixels. In the output layer, the defuzifieation process is done and the restored image is obtained. A learning method based on the advanced genetic algorithm is adopted to adjust the network parameters from a set of training data. Experimental result shows that the FNN filter is effective in noise canceling and simple to be realized by hardware.