Blemish Removal from Images using Edge Detection and Gaussian

Gregory Yoblin and Joe Marino

Smoothening When a person goes in to get their photograph taken, they do
not always look their best.  Pimples show up overnight, food is stuck to
teeth, and skin can be overly greasy, among other problems.  Currently,
there exist methods for manually removing such blemishes from pictures. 
However, these take time, patience and skill, and can still be
incomplete (people do make mistakes).  What we propose is to create a
program that is capable of searching out and eliminating all blemishes
from the picture.  This will be done using a variety of image processing
and machine vision techniques including, but not limited to: Canny based
Edge Detection, Connected Components, Region Detection and Removal, and
Gaussian Smoothening. 

This project will be done incrementally, in order to ensure that the
process is working at all phases of development.  Our first phase will
be to manually select blemishes, remove them by hand, and use Gaussian
blurring to eliminate the "holes" left by the removed blemish.  The next
phase will involve region detection and removal.  Beginning with a
pre-selected frame, edge detection will be run (or simulated) and a
"dot" will be placed at the center of gravity of the edges.  This dot
will be used for region detection.  The region will then be maxed out
(255) and then blurred over using the algorithm from the previous phase. 
The next phase will involve implementing a Connected Components
algorithm to determine if edges found are to truly be considered a
blemish.  This will be done to hopefully eliminate the side of the face
from blemish consideration.  The final and most complicated phase will
involve the base edge detection and density calculation to determine
automatically if something is a candidate for "blemish-hood."

Decisions on the methods to be used were based on a number of resources. 
Team member Greg Yoblin spent the most recent summer as a research
assistant at the University of Central Florida working in the field of
computer vision.  There, Professor Niels da Vitoria Lobo taught a 3-week
course on the fundamentals of Computer Vision, including Roberts, Canny,
Sobel, and LoG (Mexican hat) Operations.  Of these methods, Professor
Niels da Vitoria Lobo pointed out that for basic work, the most
functional method was Canny Edge Detection.  Also, the canny algorithm
is used as the baseline for comparison with new edge detection images
(1).  Based on this, we decided that it would be an appropriate method
to use.  The choice of Gaussian blurring was based on several sources
showing that elementary blemish removal is done by painting on a
Gaussian blur, a technique which remains a part of more comprehensive
efforts (2,3).  Additionally, the Edge Density detection will be done
based on the work of researchers at Michigan State University (4). 

REFERENCES

(1) "Comparison of Edge Detectors: A Methodology and Initial Study"
(Heath, et al) (CVPR, 1996)

(2) http://graphicssoft.about.com/cs/photoshop/ht/apsenhanceskin.n.htm

(3) http://www.pcworld.com/howto/article/0,aid,119116,00.asp

(4) "What the Visual System Sees": The relationship between fixation
positions and image properties during a search task in real-world scenes
(Mack, et al)
(http://people.vanderbilt.edu/~michael.mack/docs/OPAM2003.pdf)