An exploration of thresholds in retinex theory towards digital image processing of remotely sensed imagery

Optical remote sensing employs radiometric corrections to obtain measures of surface reflectance. Depending upon the particular application, these procedures may require intensive image processing and the specification of many external input parameters. This study uses principles derived from human colour vision research to help design new radiometric correction procedures that are simple, image-based, and perform reliably for a variety of imaging conditions.; The research is based on Edwin Land's retinex theory. Land used decision rules, or thresholding procedures, in an attempt to restore relative reflectance from scenes under conditions of varying electromagnetic flux (i.e. , due to uneven illumination). Several algorithms, derived from retinex theory, were designed such that they differed in the decision rules used to distinguish reflectance changes from illumination changes in test images. Global threshold procedures were employed in two of the algorithms and a third was based on locally adaptive thresholds derived from measures of image texture. Different images, ranging from simple artificial images to real world scenes, were used to test the algorithms' performance.; The global algorithms succeeded in the removal of uneven illumination with artificial images. The same algorithms were then tested on remotely sensed images exhibiting various degrees of effects analogous to uneven illumination. The uneven illumination effects remained in the images after processing. It was found that the global algorithms had difficulty coping with the presence of image texture. Local texture differences rendered the global thresholding procedures ineffective. This research demonstrates that locally based threshold adjustment techniques are superior in dealing with texture than the global thresholding Land employed.; Decision rules as devised by Land are too simple for real world settings, because they do not consider image texture. A new conceptual framework for retinex algorithms, based on locally adaptive thresholding, is hypothesized.