150 likes | 233 Views
COMPLEX SPORT IMAGE CLASSIFICATION USING SPATIAL COLOR and POSTURE CONTEXT DESCRIPTORS and NEURAL CLASSIFIERS. P.Panakarn*, S.Phimoitares, and C.Lursinsap Advanced Virtual and Intelligent Computing ( AV IC) Research Center Department of Mathematics,C hulalongkorn University, Bangkok,Thailand.
E N D
COMPLEX SPORT IMAGE CLASSIFICATION USING SPATIAL COLORand POSTURE CONTEXT DESCRIPTORS and NEURAL CLASSIFIERS P.Panakarn*, S.Phimoitares, and C.Lursinsap Advanced Virtual and Intelligent Computing ( AV IC) Research Center Department of Mathematics,C hulalongkorn University, Bangkok,Thailand
GOAL • Improvement of image searching. • To know what sport is in the image. • Find features for good classification accuracy. • More descriptive of the postures.
outline • Other features • Majority color extraction(color histogram) • Descriptor for complex background(DCT) • Posture descriptor(Cb,Cr) • SVD on DCT,Cb,Cr • Experiment
Other features • Edge histogram • Region-based shape • EH & RS will compare with the feasure proposed later.
Majority color extraction • RGB color 64bin colors • Use the most significant two bits of each color channel. • Make histogram
Descriptor for complex background • Change from color domain to frequency domain • Discrete cosine transform • RGBgrayDCT
Posture descriptor • RGB YCbCr • No Y because it is too sensitive to colors
SVD on DCT,Cb,Cr • The three matrices , DCT,Cb,Cr is the image size. • The essential information must be extracted. • Diagonal elements will be used.
SVD on DCT,Cb,Cr • For DCT,Cb,Cr matrices • SVD(single valued decomposition)
Experiment • 300 images with 6 sports each • Baseball, Basketball, Field and Track Skiing, Soccer, and Swimming • 200 for training , 100 for testing • 130*200 pixels
Experiment • The elements after SVD is 130 for DCT,Cb,Cr matrices • Features are 130*3+64=454 features • Compare with (EH & RS) using NNC,RBF