We are mostly aware of color photography. The Red, Green and the Blue channels are just three frequencies of the spectrum. In multi-spectral imaging other frequencies of the spectrum are acquired. Such multi-spectral imaging finds applications in a variety of scientific problems in agriculture, surveillance, military etc.

However, multi-spectral cameras are extremely expensive. This owes to the cost of the sensors - within the optical range, the sensors are cheap; this is the reason why digital photo cameras are so inexpensive. But, once outside the visible range of the spectrum, the cost of sensing sky rockets. On top of that, multi-spectral cameras use full sensor arrays for each spectral channel. This not only increases the cost, but also makes the acquisition device bulky with movable mechanical parts. In turn, this results in pixel registration problems across the channels.

Our work is influenced by the success of single sensor architecture of commercial RGB cameras. At each pixel location only one of the channels will be acquired, the unsampled locations will be interpolated using some demosaicing algorithm. However the famous Bayer pattern cannot be used directly - since it is tailored for the visible range. We propose an uniform sampling array in pseudo-colors.

Linear interpolation is carried out by learning the interpolation weights. The results are encouraging.

We are getting very accurate interpolation results for even 5 bands, i.e. with just 20% sampling!. Below are the results on 5-band multi-spectral images.

Original Image

Undersampled Image

Reconstructed Image

The sampling pattern can also be random (but equal sampling for all channels). In such a case, the sensor array would like

Research on other areas of multi-spectral imaging are also pursued. We are looking at the possibility of recovering hyper-spectral images acquired by Rice Single-Pixel Camera. Also, we are working on denoising problems in multi-spectral imaging.