画像のスタイル特徴量 st 2nd LBP pattern index 3rd.5,rotation invariance,uniformity: #(+ ) < 5 Saturation Value 局所 特徴量の導入 HSV feature Hue 2 48/:8/8>9>3=4D9-69.5= 視覚辞書に基づく手法の比較 visual word feature vector 2 4 5 4=>92<,79! % 4=>92<,79! % 4=>92<,79 ') 8=<2?6,<=,7:6482969.,6,>?<= #9=,648.C/>.>498645 8=' ( with bit changes > 5 LBP (28 dim.) 8 bins Concatenate Saturation 5 bins HSV Value 8 bins (36 dim.) ROI 3 '><95*4/>3 (<,8=9<7 4=>92<,79.969<,8/2<,C=.,6 スタイル特徴の標本化と辞書化 Local histograms can be Generate histogram 3 4 () efficiently computed with 2 7 () 2 integral Omit binaryimages codes LBP feature ",8C=>,>4=>4.,6,>><4-?>=9 '! 478=498 K K K </?.>498 A4>3#$><,48482 A4>3><,48482/,>, 969< Hue '><95 (B>?< V cluster 28 patterns HSV S A similarity measure for illustration style, E. Garces, et. al. ACM TOG 24 '3,/482 image pyramid H 既存手法との特徴量比較,,,,,,, 256 patterns.5 LBP 34=>92<,79>3 >9, A9</ 4.>498,<C9=>C6A9</=,>?<.6?=><482-C "" ",534=>92<,79=>C6 A9</E=9..?<<8.= スタイル識別子の計算手順 An Unsupervised Approach for Comparing Styles of Illustrations, T. Furuya, S. Kuriyama, R. Ohbuchi, CBMI 25, Best paper award BoV = [,,.2,.5, ] # clusters 34=>92<,79>37,8,8/ @,<4,8. >9 A9</= FV = [.,É.2,.,É.3, É] #clusters (2#feature dimension) Aggregate Extract Generate Block of 25x25 by FV local features image pyramid pixels moved at Region-Ofevery 8 pixels Interest (ROI)Blockwise.4.5 features are.68 efficiently computable A set of LBP FV-LBP via integral images.5.22.8 Illustration Image pyramid Concatenate vectors A set of HSV FV-HSV.4.5.5.22 Fused Fisher Vector (FFV) (~2K dim.)

イラスト特徴量における 疎性 の解消 ２段階での特徴次元数の削減 %9A<! 89<7,64D,>498.,8<79@ =:,<=4>C9>34=3<@.>9<= kernel KPCA with dot (linear) v = v.3 / v LBP FV LE (non-linear) Feature manifold v = v. / v FFV feature space (e.g. ~2K dim.) KPCA subspace (e.g. 52 dim.) HSV FV Principal Component Analysis (PCA) vs. Kernel PCA (KPCA) LE subspace (e.g. 28 dim.) 多様体学習 T : # of training illustrations in database. D : # of dimensions for aggregated visual feature. PCA EVD KPCA with dot kernel Covariance matrix (DD) EVD ~2, ~2, Kernel matrix (TT) &,A/,>,!,:6,.4,8 4287,:= % 4,59 4,59 Faster to compute. 性能評価のためのデータ集合 検索精度比較結果 (/3) =,7=>C66,-6 Query (3=,7/,>,=>.98/4>498=,= >3+,<,29D, /9-E=7>39/ Style_6 Zaragoza & Adobe Style_8 Similar style (thick black contour, bright color) Style_254 Style_292 Style_266 Style_266 47,2=9:<.,>29<4D/ =>C6= donated by Microsoft corp. D-SIFT FV LBP+HSV FV (Our method) 47,2=9?8589A8=>C6= A34.3,<?=/,= extracted from ArtExplosion dataset

検索精度比較結果 (2/3) 検索精度比較結果 (3/3) =,7=>C66,-6 Query Zaragoza & Adobe =,7=>C66,-6 Query ry Zaragoza & Adobe Similar style (two-color, no shading) Similar style (color, shading) D-SIFT FV D-SIFT FV LBP+HSV FV (Our method) LBP+HSV FV (O (Our method)) map 平均適合率 による性能評価 Query Precision / 2/3 Average avg. Precision.8 3/4 Query2 Precision / Mean avg. Average Precision Average avg. Precision.83 2/3 Local feature FV-LBP.45.45.4.35 map.4.35.3.25.5.3. FV-LBP BF-LBP 2 3 4 5 Aggregated feature dimension 3 4 5 6 KPCA+LE > KPCA only > no compression.7.65 peak =.62.55 FFV (8,432dims.) FFV (KPCA) FFV (KPCA+LE).4.35. FV-HSV BF-HSV.5 2 3 4 5 Aggregated feature dimension Best for artistic paintings =.47.5.2 2 最終性能.45.5.5.6.25.2 38.6 map [%] map.5 37.5 9< ') 969<.5 44.9 Color + Shading + Texture + Stroke [Zaragoza & Adobe] 9<! % (B>?< 54.8 FV-HSV Global feature Fisher vector (FV) vs. Bag-of-features (BF) map FFV (FV-LBP + FV-HSV).82 (82%) 局所特徴量化による性能向上.3 2 4 6 8 Number of dimensions 478=498,6</?.>498 >3:<9<7,8. %,5:<9<7,8. 4=9->,48/,> /478=498!,:6,.4,8 4287,:=?..==?66C648,<4D>3,>?<=:,.

描画特徴に基づく画像検索システム 結論と考察 '>C6,>?<=H 98>8>,>?<= ",5>37=47:6 (,5.,<=:,<=4>C '>C6I 48 2<,48/ 98>8> '>C6<><4@,6H 98>8><><4@,6 ==8>4,66CB:69<,>4@ %<74>.97:<974= &.,66 %<.4=498 8><,...?<,.C 受賞歴 今後の課題 (/2) 4<.>,-6B:69<,>4989=>C6=:,. '>C6/:8/=98,::64.,>498=,8/.?6>?<!482?4=>4.6,-6:<9:,2,>49898/7,8/ %<92<==4@48><,.>498A4>348>?4>4@.?= 今後の課題 (2/2) '>C6,>?<=4828<,6 <97/<,A482=>97,8?,.>?<= <9==/97,48.98=4=>8.C C-<4/,8,6C=4= 2 A4>3@.>9<9< /,>, ご静聴 ありがとうございました