NASA / Langley Research Center / Electromagnetics and Sensors Research Branch

Retinex Image Processing

small logo

 


Background


RETINEX IMAGE PROCESSING:IMPROVING THE VISUAL REALISM OF COLOR IMAGES

A versatile automatic method for enhancing images....

The goal of our work is to make the recorded image look like what you would have seen if you had been observing the scene in person. This means full visual realism analogous to the "concert hall" sound of audio fidelity.

What's Wrong with Pictures?

Recorded color images suffer from significant losses in visual quality (compared to the direct observation of the scene by the eye) when there are either spatial or spectral variations in illumination. The visibility of color and detail in shadows is quite poor for recorded images and a spectral shift in illumination either toward the blue or the red reduces the overall visibility of scene detail and color. These lighting defects are quite common. Likewise for scenes with some white surfaces (clouds or snow, for example), visibility of color and detail in the non -white zones of the image is poor. Therefore a general purpose automatic computation is needed to routinely improve images. A dramatic example of this kind of image defect is shown below.

original image processed image

We have developed such a computation which synthesizes dynamic range compression, color constancy,image sharpening,and color rendition. Our starting point was the last retinex concept proposed by Edwin Land which has a center/surround spatial structure akin to the receptive fields of the neurophysiology of vision. Initial experiments yielded promising results as shown below.

color constancy examples

 

 

 

In order to develop this concept to the level of a practical processing scheme, we had to 1) resolve fundamental design issues for the center/surround retinex, 2) extend the single scale design to a multiple scale design, and 3) develop a color restoration which overcomes the gray-world assumption intrinsic to retinex processing. All three of these elements were necessary for the development of a practical computation- the Multi-Scale Retinex with Color Restoration (MSRCR ). Further the specification of the computation was strongly guided by the comparison of experimental results to THE DIRECT OBSERVATION OF TEST SCENES. Only in this way could we be certain that the computation was performing in a functionally similar manner to human vision. Our working design has now been extensively tested on a variety of test scenes and hundreds of test images. We have included many space images in our testing to demonstrate the advantages of the computation for remote sensing images and Space Shuttle imagery.

Since the technique is also useful for consumer color images we show its performance for a variety of everyday images.

These examples provide a varied illustration of the benefits of dynamic range compression including the need for it in wide dynamic range reflectance images, however the need for color constancy as a very practical image enhancement requires further illustration as described below.

original image processed image original image processed image

The remote sensing image is strongly "blue" probably because of high concentrations of atmospheric water vapor while the construction site is shifted toward the red perhaps because of a very low sun angle. In both images the visibility is poor and cannot be improved simply by advancing the gain of the image. By comparison the MSRCR improves the visibility of the overall scene detail by reducing the color shift in the images (as well as automatically performing the dynamic range compression).

While we are still in the preliminary phase of performance comparisons of the MSRCR with other image enhancement methods, we have examined this retinex relative to manual gain/gamma/offset corrections, automatic histogram equalization, manual "burning and dodging", and homomorphic image processing

Overall, the retinex performs automatically and consistently superior to any of the other methods. While the other methods may work well on occasion cooperative images, it easy to find images where they perform poorly. The retinex performs well on cases where the other two methods are clearly not appropriate.

NOTE: All color images have been processed with exactly the same computation. This provides strong evidence for the generality of the computation and its practical value as an automatic means for enhancing images. There was no change in constants or any other alteration or auto-gain applied to these images.

The MSRCR is available for licensing to commercial partners.

 

Responsible NASA Official: Glenn Woodell