Abstract

Visualization of high dimensional data such as hyperspectral data is a key step for analysis. A recurrent issue is to represent the data as a RGB image for easy display. This compression is usually obtained using priors on the data or unsupervised meaningful information extraction which may result in information loss. Rather than making any assumption on the type of data, we propose to compress the data into a three channel image such that it is possible to reconstruct the original . For this, we use deep convolutional autoencoders in the spectrum dimension. We also show that our approach can be extended to take external knowledge into account such as visual hints for color display.