Straight-Edge Extraction in Distorted Images Using Gradient Correction

Many camera lenses, particularly low-cost or wide-angle lenses, can cause significant image distortion. This means that features extracted naively from such images will be incorrect. A traditional approach to dealing with this problem is to digitally rectify the image to correct the distortion, and then to apply computer vision processing to the corrected image. However, this is relatively expensive computationally, and can introduce additional interpolation errors. We propose instead to apply processing directly to the distorted image from the camera, modifying whatever algorithm is used to correct for the distortion during processing, without a separate rectification pass. In this paper we demonstrate the effectiveness of this approach using the particular classic problem of gradient-based extraction of straight edges. We propose a modification of the Burns line extractor that works on a distorted image by correcting the gradients on the fly using the chain rule, and correcting the pixel positions during the line-fitting stage. Experimental results on both real and synthetic images under varying distortion and noise show that our gradient-correction technique can obtain approximately a 50% reduction in computation time for straight-edge extraction, with a modest improvement in accuracy under most conditions.