Techniques are provided for detecting edges in a color image by analyzing variations in multiple color channels. No greyscale conversion is performed, thereby preserving color differences which might be lost in greyscale conversion. Gradients in multiple color channels are merged to compute a single edge snapping vector field from which edge locations are derived. By retaining the multidimensional characterization of the colors that comprise the processed image, a maximum color difference can be calculated using parameters which closely correspond to human perceptions of color. Maximum color difference is formulated using an LUV color model. The Euclidian distance between two pixels characterized using a multidimensional color model more closely approximates a perceived color difference between as compared to the Euclidian distance between the same pixels characterized using a greyscale color model. An edge snapping vector field is derived from these Euclidian distances, thereby providing a more accurate basis for edge detection.