An example of tuned neural network based noise reduction filters for images

This paper presents some results on noise reduction in digital images using artificial neural networks. The design is based on the known capacity of supervised neural networks to learn from examples, avoiding the need for explicit knowledge about the image distortion function. The filter is implemented using current backpropagation feedforward neural networks, and works on the first differences calculated between neighbor pixels. The filtered gray level images are obtained from the output of the filter using an iterative reconstruction algorithm. We give some experimental results which show that the neural network filter provides an increased reduction in noise variance, when compared to the median filters