A Neural Network Approach to Classifying Cartoons Based on ColorProject PlansPowerPoint PresentationProject StepsSlide 5Data CalculationsSlide 7ResultsSlide 9Slide 10By: Jared MeyerA Neural Network Approach to Classifying Cartoons Based on ColorECE 539 Final ProjectProject Plans•Collect Data–Choose 20 different animated series–Choose 3-5 episodes at random•Varying seasons if applicable–Covert each episode to series of images•Calculate data for each image–Write program in C#•Build Artificial Neural Network with bp.m–Find structure/data combination that maximize classification rate•Why?–Big fan of cartoons/animated series–Interested in how images are represented in computers–Would be neat to see color patterns in shows•Existing Results–Weather classification based on color (Moosmann, 2008)–Linear kernel Support Vector Machine–3 classes•Clear•Light Rain•Heavy Rain–Average classification rate: 89%Project Steps•Ripped 3-5 episodes of following shows:Avatar: The Last Airbender The Real Ghost BustersBatman RebootCourage the Cowardly Dog Samurai JackCowboy Bebop The SimpsonsEd, Edd, n’ Eddy South ParkFamily Guy SpidermanFuturama Spongebob SquarepantsInvader Zim SupermanOutlaw Star Teenage Mutant Ninja TurtlesPowerpuff Girls Teen Titans These form the 20 outputs for ANNProject Steps•Converted episodes to images–X Video Converter–One BMP image per 200 frames•Remove first frame–Usually pure black•Remove all frames including end credits–Would add biasData Calculations•Wrote program in C# to calculate 14 Features per image•Brightness, Contrast, Saturation, RGB ratio•‘Lininess’–Pixels with large brightness differenceData Calculations•‘Important Areas’–Pixels brighter than average brightness•Counted Red, Orange, Yellow, Green, Blue, Violet, Grey Pixels in ‘Important Areas’•Finally, used bp.m program to build ANN using back-propagation algorithmResults•Data varied greatly, even in same episode–~5% classification rates•Averaged 10 random frames togetherResults•Contrast, Color counts still varied too much–Removed them; didn’t show much patternResults•Much better classification rates with new dataFinal Results: 5 Features: Brightness, Saturation, RGBANN Structure: 2 Hidden layers, 9 neuronsClassification Rates: 57.14% on Training 47.50% on TestingPretty good, considering we had 20
View Full Document