DOC PREVIEW
UCSD CSE 190 - Nasolabial Wrinkle Classification

This preview shows page 1-2 out of 6 pages.

Save
View full document
View full document
Premium Document
Do you want full access? Go Premium and unlock all 6 pages.
Access to all documents
Download any document
Ad free experience
View full document
Premium Document
Do you want full access? Go Premium and unlock all 6 pages.
Access to all documents
Download any document
Ad free experience
Premium Document
Do you want full access? Go Premium and unlock all 6 pages.
Access to all documents
Download any document
Ad free experience

Unformatted text preview:

Nasolabial Wrinkle Classification David Lluncor University of California, San Diego 9450 Gilman Drive, La Jolla, CA, 92092 [email protected] Abstract Categorization of facial wrinkles is used in dermatology studies to determine the efficacy of a certain treatment over a certain period of time. These studies, however, are limited by the fact that when dermatologists assign these categorizations, it is done so through qualitative, in-vivo observation. The objective of this study was to train a computer to categorize nasolabial wrinkles into one of four categories based on one image of the patients face. Preliminary results suggest that the computer will be able to correctly categorize category 1, 2, 3, and 4 wrinkles with 88.2%, 57.3%, 26.2%, and 80% success, respectively. 1. Introduction As the fields of cosmetic surgery, dermatology, and plastic surgery continue to grow, more and more products are developed to help eliminate blemishes and facial wrinkles. Dermatologists have developed several methods [1] to categorize wrinkles into distinct categories depending on wrinkle length, depth, and overall how it looks. There are several methods which take into account different aspects of a wrinkle, but all in all it is up to the subjective observation of a dermatologist to assign these wrinkle categories. While there is a fairly strong correlation among dermatologists who assess the same patients [2], there is little quantitative evidence that is used in the process of determining wrinkle severity. Companies who develop aesthetic products use these scores to determine how well a treatment has helped to alleviate wrinkles over a period of time. Some have quantified wrinkles using plastic molds [2] to measure the actual depth of a wrinkle and using its depth to categorize the wrinkle. This technique is quite invasive and our goal is to create a method which is cheap, effective, and has a quick turnaround time. In this paper I develop an approach to categorize wrinkles based on various light intensities that emit from the skin, where a more pronounced wrinkle is likely to have less light intensity emission at the point of the wrinkle. The preliminary results (n = 23) show that indeed the “intensity profile”, that is the pattern of light emission, from wrinkles of category 1-4 exhibit distinguishable patterns. 2. Approach It is presumed that from measuring the light intensity values across a part of the face which contains the wrinkle we canextrapolate the physical characteristic of how deep a wrinkle is. While we are not actually measuring the physical depth of the wrinkle itself, we are using the fact that a more pronounced wrinkle produces a greater shadow to leverage taking solely the intensity profiles from an image. The images were taken using a Visia Camera. The method for extracting the intensity profile from an image can be broken into three simple steps. 2.1 Visia Scanner Fig 1. The Visia scanner was used to taken images from 45°, 0°, and -45° respective to the front of the face in seven different illumination settings. The Visia skin analysis machine is a standard in skin complexion analysis [3]. The scanner can take pictures using up to seven different illuminations, however, there was only one illumination used from the 0° perspective to analyze the data. In future studies, I may consider using different perspectives and illuminations to analyze the data. 2.2 Relevant Pixel Markers A standard approach was used to extract four points on the face: the right corner of the lip, the left corner of the lip, the right corner of the nose, and the left corner of the nose. Fig 2. Example of where the four points that were selected from each patient. These four markers were selected because they can easily be detected by certain face recognition algorithms [4]. In future works the process will be automated such that rather than a user inputting where to select relevant pixels, a computer will ideally locate where the corners of the mouth and nose are. 2.3 Intensity Profile Lines Using the four point markers, a line was then extracted from the black and white image of the light intensity values. Two different approaches were used to take the intensity profile: averaging values and the worst-case value.Fig 3. Example of where the intensity values were taken from when analyzing the worst-case. The worst-case line, selected among 10 parallel lines, is defined as the line which had the greatest difference in lowest and highest intensity value within the line consisting of approximately 80 pixels. Fig 4. Example of where the intensity values were taken from when the average-case. In this approach, all lines were treated equally and the intensity values were averaged at concurrent locations in each of the lines. 3. Training and Testing The k-nearest neighbor approach was used to categorize the wrinkle based on two input parameters: the the fit slope (m) and fit b-intercept (b) of the intensity values to the line y = |x| for a particular wrinkle. Majority voting was used when k = 3 and k = 5 to determine the class, while ties were broken based on the class of the nearest neighbor. Neighbors were determined using the Euclidian distance between the two input parameters. There were 17 intensity profile examples from category one, 11 intensity profile examples from category two, 12 intensity profile examples from category 3, and 6 intensity profile examples from category 4. To use cross validation, one intensity profile was omitted from each of the four categories and was used as the test data while the rest of the examples were used as training data. To fully support the use of five-fold cross validation, the number of examples from category 4 will need to increase to match the number of examples provided by categories 1, 2, and 3. Also, less data will need to be used as training data in each configuration. As a result of such a method for cross validation, there were 11,220 different configurations for which the test data and the training data was chosen.4. Data and Results There were 23 patients whose nasolabial wrinkles were analyzed. There were 26 patients in total whose intensity profiles were collected, but because of mustaches and other visual distractions that would hinder the use of the algorithm described in this paper, those 3 patients were omitted. Fig 5. Graph of the intensity values across 23 patients taken from the worst-case. Based on the


View Full Document

UCSD CSE 190 - Nasolabial Wrinkle Classification

Documents in this Course
Tripwire

Tripwire

18 pages

Lecture

Lecture

36 pages

Load more
Download Nasolabial Wrinkle Classification
Our administrator received your request to download this document. We will send you the file to your email shortly.
Loading Unlocking...
Login

Join to view Nasolabial Wrinkle Classification and access 3M+ class-specific study document.

or
We will never post anything without your permission.
Don't have an account?
Sign Up

Join to view Nasolabial Wrinkle Classification 2 2 and access 3M+ class-specific study document.

or

By creating an account you agree to our Privacy Policy and Terms Of Use

Already a member?