Detecting edges with sub-pixel precision in JPEG images

Edge localization is one of the mostly used procedures in image processing. In some applications it is desirable to improve edge detector precision to sub-pixel values. The input data for precise edge detector are usually in the form of uncompressed image. Sometimes, only JPEG images are available for processing. In our paper we deal with this situation, we focus on the question how lossy JPEG compression of image can change the precision of edge detection. We used JPEG images with different quality to test three methods for edge localization with accuracy at the sub-pixel level for 1-D images. First two algorithms are based on spatial and gray level moments of the image function. Third method used in our experiments approximates samples of real image with erf function. To compare these three methods we used noiseless and noisy JPEG images, two types of noise were used: Gaussian noise and `salt & pepper'. The standard deviation of edge position error was chosen as the precision criterion.