|
• Image Compression via Neural Networks
One way of compressing data using neural networks involves construction of a 5-layer bottleneck network. This neural network is an ordinary feedforward neural network, with inputs and outputs the same (the neural network tries to predict the inputs), and a middle layer which is significantly smaller than the number of inputs. The idea is to have the neural network reproduce the input values at the output nodes, but force the information through the small bottleneck to get it to comrpess.
This could be applied to image compression by having the neural network read, say, a 4x4 square (=16 variables) of pixel values and try to repreoduce them as its output. If the middle layer is limited to, for instance, 8 nodes, then a compression of 50% is achieved (8 being 50% of 16). During compression, the neural network would be fired halfway, and the result of the 8 hidden nodes would be stored. Decompression involves firing the second half of the neural network.
This is a form of "nonlinear principal components analysis".
-Will Dwinnell
http://will.dwinnell.com
|
