Prediction of wavelet transform coefficients using neural networks applied to lossless compression of multispectral images

摘要

We present a lossless compressor for multispectral images that combines two classical tools: wavelets and neural networks. Due to their huge dimensions, images are split into small blocks and the wavelet transform that maps integers to integers is applied to each block -and each band- to decorrelate it. In order to increase even more the compression rates achieved by the wavelet transform, coefficients in the two finest scales are predicted by means of neural networks, which use causal information (ie, coefficients already coded) to get nonlinear estimates. In this work, we add coefficients from other spectral bands to compute the prediction, besides those coefficients belonging to the same band, which lie in a causal neighbourhood. The differences are then coded with a context based arithmetic coder. Several options regarding initialization, training and architecture of the neural networks are analyzed. Comparison results with other lossless compressors (with respect to the coding time and the bitrates achieved) are given.