Edge detection – An unexplored byproduct of TIE

As a part of my Doctoral work, I developed a method to enhance quality of interpolated images at the cost of a little additional data. We named the method Targeted Image Enhancement (TIE). An unintended but interesting by-product of the targeting algorithm in out method was that it threw up images that looked very much like the output of an edge detection algorithm. Here is a link to our paper on TIE.

We evaluated other popular edge detection techniques as alternatives to our method in our pipeline. For our test cases, our method proved to be the best and so we retained it. However, this also threw up the question of how do we compare it with other edge detection techniques. This in turn led to the question of how the results (not necessarily the computation complexity) can different edge detection methods be compared.

A few examples of the intermediate product of TIE that can be compared to edge detectors are shown below. This involves a threshold. By changing the the threshold value, the number of edge pixels detected change.

Edge detection using TIE in the RGB space.
Edge detection using TIE

An example of comparison with Canny edge detection

The two images show edges as detected by Canny and TIE. The number of edge pixels are highly dependent on thresholds chosen. These two images were created by using thresholds which gave a comparable compression in our Targeted Image Enhancement (TIE) method. The thresholds were not tuned for visual edge quality.

Edge detection using Canny
Edges in Monarch per Canny with thresholds of 150 and 50.

Edge detection in TIEEdges in Monarch per TIE at a threshold of 24.

If you found this post interesting, you might like the following

A Non-Linear Method to Interpolate Binary Images Using Location and Neighborhood Adaptive Rules

A Yahoo Videos story

A patent filing for an invention by the author on measuring internet video quality.