Although edge detection has been the subject of extensive work for many years, it is still not difficult to find or manufacture images where the edge picture produced by current methods is visually unsatisfying. Human observers have a sophisticated ability to determine which intensity transitions in a scene are caused by edges of physical objects or zones, as distinguished from texture, illumination changes, and other artifacts. Apparently, our existing algorithms do not reflect the characteristics of the human visual system (HVS).
The authors skillfully pursue this concept, defining a method for identifying edges that are considered significant by the HVS. Beginning with the lifting scheme, a discrete wavelet transform, they find areas of significant contrast. Incidentally, they describe a complementary approach for finding areas without visually significant edges; this edge-avoiding framework is useful in its own right.
Next, they incorporate the most innovative part of their approach: emulations of certain characteristics of the HVS. Variations in contrast sensitivity, due to intensity level and spatial frequency, as well as visual masking--the suppression of edges in visually busy regions--are modeled by adjusting the contrast detection equations. The implementation of these methods necessarily involves a number of parameters, and the authors conduct a well-documented set of experiments to calibrate the model.
The resulting technique is demonstrated in three application areas. First, image resizing by intelligent seam carving is performed using the map of visually significant edges to select the seam locations; performance on a single test image is good. Second, tone remapping of an image based on the HVS model achieves impressive results. Finally, the model is used to augment Ward’s technique for stitching images together into a single panorama [1].
This well-written paper presents an intriguing framework for determining the visual significance of edges within a scene. Since the method emulates the HVS, it will find utility in applications that produce imagery for human viewing, as the applications described in the paper show. However, it is not necessarily true that the most significant edges to a human observer are always the most significant in a purely automated context--for example, inspection--and some care should be taken in the application of this method to those areas.