Images Faster and Better

Demo 1 was created by taking the original 512x512 Lena IEEE test image, and its resolution was reduced by 4x by resampling (average method on each 4x4 pixels block "box convolution kernel"), creating a 128x128 image. This reduction method models sensors of square size collecting the light evenly (such as Foveon X3© sensors) 3. Other sensor characteristics (such as Bayer with regards to each color plane) can also be assumed as a-priori knowledge of how the image was acquired, without changes to the speed of the reconstruction. This makes this algorithm ideal for a combined demosaicing and 2X zooming in one pass.

The image was then reconstructed using bicubic, Lanczos filters, the "Backprojected Jensen-Xin Li" interpolation and the new Comagna technology for comparison.

Backprojected Jensen-Xin Li was selected for this comparison as it has been found as having one of the best signal-to-noise ratio among a number of methods, while the only, marginally better methods were slower by yet another order of magnitude.1

Demo 1 (Comparison summary) PowerPoint Presentation

Demo 2 In this comparison, we blow up a 64x64 area of the Lena image to 16x in each dimension, that is to 256x the size of the image, again, using bicubic, Lanczos and Comagna techniques. Again, box kernel characteristics of the ideal square sensors are assumed for the image acquisition.

Demo 2

Demo 3 This demo is a variation of Demo 1, however the resampled image was also compressed using JPG compression. The artifacts created by the compression are emphasized by the reconstruction, however Comagna method is designed to be far less sensitive to these artifacts. Two variations of the Comagna method are shown, the second being even less sensitive to artifacts.

See Demo 3

Demo 4 shows a very different application of the basic Comagna technology. In this demo the original Lena image was compressed drastically using JPG. The compression shows major artifacts. Comagna technology was used to restore/improve the JPG image. The result is better both visually and numerically (as opposed to filtering, smoothing methods). The artifacts are removed or lessened without losing details of the original image. 

See Demo 4


Foveon X3 sensor, Wikipedia