Lossless Multi-component Image Compression Based on Integer Wavelet Coefficient Prediction using Convolutional Neural Networks

This work extends the single component Wavelet Prediction Compression (WPC) framework proposed earlier to lossless compression of multi-component imagery. WPC employs a series of convolutional neural networks (CNNs) to generate predictions of wavelet detail subbands, using subbands within the same discrete wavelet transform (DWT) decomposition level. Prediction residuals are then coded in place of original coefficients in the compressed codestream. At the decoder, predictions are reproduced using an identical set of CNNs, and combined with residuals to achieve perfect subband (and image) reconstruction. The proposed Multi-Component Wavelet Prediction Compression (MCWPC) framework extends WPC to multi-component by considering both inter-and intra-component redundancies, resulting in improved spatial and spectral decorrelation of DWT coefficients. Compression results (Table 1) show MCWPC achieves 7.2% and 23.8% bit-rate reductions over JPEG2000 in the YUV and RGB colorspaces, respectively.