Image dequantization for hyperspectral lossy compression with convolutional neural networks

Significant work has been devoted to methods based on predictive coding for onboard compression of hyperspectral images. This is supported by the new CCSDS 123.0-B-2 recommendation for lossless and near-lossless compression. While lossless compression can achieve high throughput, it can only achieve limited compression ratios. The introduction of a quantizer and local decoder in the prediction loop allows to implement lossy compression with good rate-performance. However, the need to have a locally decoded version of a causal neighborhood of the current pixel under coding is a significant limiting factor in the throughput such encoder can achieve. In this work, we study the rate-distortion performance of a significantly simpler and faster onboard compressor based on prequantizing the pixels of the hyperspectral image and applying a lossless compressor (such as the lossless CCSDS CCSDS 123.0-B-2) to the quantized pixels. While this is suboptimal in terms of rate-distortion performance compared to having an in-loop quantizer, we compensate the lower quality with an on-ground post-processor based on modeling the distortion residual with a convolutional neural network. The task of the neural network is to learn the statistics of the quantization error and apply a dequantization model to restore the image.