Introduction Collecting data Pixel analysis Training data Results conclusion Questions/live demo Android App for Identifying Digital Signage Viewer Dane Hylton Kennesaw State University April 27, 2017 Dane Hylton Android App for Identifying Digital Signage Viewer
21
Embed
Android App for Identifying Digital Signage Viewerksuweb.kennesaw.edu/~mkang9/teaching/CS7455/548553-1051953 … · Introduction Collecting data Pixel analysis Training data Results
This document is posted to help you gain knowledge. Please leave a comment to let me know what you think about it! Share it to your friends and learn new things together.
Transcript
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Android App for Identifying Digital Signage Viewer
Dane Hylton
Kennesaw State University
April 27, 2017
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Race detection
E-Tech software predicts ethnicity based on person’s last name
Kairos: http://kairos.com/diversity-recognition
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Ethnicities used:
African/black
Asian
Indian
Hispanice/Latino
Midille Esastern
White/Caucasian
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Data collection
The data was comprised of face pictures taken from theinternet.
20 data sets of each race were collected.
Each data set included 10 male and female.
Each picture were then editted to get as close to the face aspossible (this was done to avoid background pixel data).
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Samples
Original Image Edited image
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Finding the most frequent pixels
Objective: find the most frequent pixels (frequency pixels)
Finding the most occuring pixles for each race gives and ideaof what pixel values to look for when detecting race.
Images were processed in Matlab and stored in their originaldimensions.
A matrix was formed for each race with their frequency pixels.
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Converting from RGB to HSV
HSV color
Named for 3 values: hue, saturation, and value. Why useHSV?
To analyze the pixels and sort them interms of their tone.
Will be easier to find specific range of colors.
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Sorting HSV colors
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Intervals of HSV colors for detecting race
For the 20 frequent pixels of each race the average of thebottom half (abh) and the top half was (ath).
A detected face would need to be in this range to detectsomeone’s race.
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Sample data with intervals: abh tbh
For the 20 frequent pixels of each race the average of thebottom half (abh) and the top half was (ath).
A detected face would need to be in this range to detectsomeone’s race.
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Sample
Sorted HSV pixels for Africanimage data
Upper and lower bound forafrican data
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Sample data with intervals: abh tbh
For the 20 frequent pixels of each race the average of thebottom half (abh) and the top half was (ath).
A detected face would need to be in this range to detectsomeone’s race.
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Neural network architecture
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Weight training
Error term for ouput
δOk = Ok(E )(1 − ok(E ))(tk(E ) − Ok(E ))
where Ok(E ) is the ouput, and (tk is for categorization.
Error term for hidden unit.
δhk = hk(E )(1 − hk(E )) ·∑
i∈outputs(wkiδOi
)
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data
Resultsconclusion
Questions/live demo
Calculating new weights
For each weight wij .
Between input unit i and and hidden unit j . The learning raten = 0.7.
∆ij = nδhj xi
Between hidden unit i and output unit j.
∆ij = nδOjhi (E )
where hi (E ) is the output from the hidden unit i for E.
Dane Hylton Android App for Identifying Digital Signage Viewer
IntroductionCollecting dataPixel analysisTraining data