Image restoration based on multiscale relationships of image structures

Aerial photographs sometimes suffer from artifacts caused by vignetting effects and changing topographic Sun-canopy-sensor geometry. We present an empirical image restoration method that is based on multiscale relationships of image structures. The fine-scale image structures depict tree crowns in a deciduous forest and serve as units in the restoration process. The color image is initially converted to the intensity, hue, saturation (IHS) system. For the I-band, two different types of variables are estimated for each segment: the local intensity difference of neighboring segments (affinity) and the mean intensity per segment. For the H- and S-bands, the mean value per segment is estimated. Regression analysis is used to model the relationship of these four variables with the coarse-scale intensity values of the corresponding segments. The correction results in new feature values that are uncorrelated with the coarse-scale intensity values. The method is evaluated on three digital aerial photographs with a ground reference dataset from the Eastern Deciduous Forest in West Virginia, USA. The image correction method is shown to result in a significant improvement for tree species classification.