1-Nearest Neighbor Filter

Single pixel regions rarely can be classified, even under the best circumstances. Therefore, we prefer to filter these out as our last enhancement stage. The 1-Nearest Neighbor filter removes single pixel outliers by replacing each pixel in the image with the mean of its value and the grey level value of an adjacent pixel which is closest to its current value. Note that one application of the 1-Nearest Neighbor filter may cause small neighborhoods of pixels to oscillate. Therefore, we apply the 1-Nearest Neighbor as an iterative filter, stopping when the input and output images are identical. For faster convergence, we use a damped approach which assigns an output pixel to the mean of its original and nearest neighbor values. Typically, we converge in roughly six to eight iterations.

Since no image enhancement occurs along the pixels of image borders, we crop the border so that additional segmentation techniques will not use this raw data to merge dissimilar regions via paths through the noisy, uncorrected pixels. For this application, we crop the border by a width of three pixels.