Denoising Based On Adaptive Neural Network Fuzzy Inference System Image

Image noise removal is a crucial content of images pre-processing, and its results have a great effect on the final results of images processing. Its main problem is how to improve the ability of noise removal with keeping the ability of preserving the details and edges of images. The major weakness of the recent image filtering algorithms is shortage of the ability of protecting the image details and edges which are smaller than the size of filtering windows. In order to eliminate this limitation, in this paper, a novel scheme for the removal of impulse noise is proposed, which has two processes:detecting noise and removing noise.In the detecting noise process, two algorithms for detecting impulse noise have been proposed. One is based on the second-order differential images, the other is a double noise detector based on the adaptive median filter and the fuzzy theory. In first method, the noise pixels are detected via constructing a first-order differential image and second-order differential image based on the fact that the difference of the noisy pixel with the nearest good pixel will be different from that of two nearby good pixels. Experimental results show that the proposed method yields a lower computational cost and higher detection performance. In the latter method, the noise candidates identified with the adaptive median filter are judged again by local fuzzy membership function to improve accurate ratio of noise detection. According to the theory of the adaptive median filter, the extreme values in the detection region will be detected as the corrupted pixels; some of them may be uncorrupted pixels. Therefore, its results were judged again by local fuzzy membership function. As a result, the detection performance was further improved.In the removing noise process, a novel method using a fuzzy inference system based on neural network is proposed. Since the fuzzy inference system not only has the ability to deal with ambiguous language information, but also can simulate human intelligence in judgments and decision making, in addition, the neural network has the self-learning, adaptive capabilities, parallel processing capabilities and fault-tolerant capabilities, the proposed method combine the superiority of the fuzzy inference system with that of neural network, such as the self-learning, adaptive capabilities, parallel processing capabilities, fault-tolerant capabilities and simulating human intelligence in judgments and decision making. Experimental results show that the proposed filter yields a better performance than the recent filters in terms of details preservation and noise suppression.