755-8611 2-16-1 HUNTEXHUNTER1 NARUTO44 Dr.SLUMP1,,, Jito Hiroki Satoru MORITA The Graduate School of Science and Engineering,Yamaguchi University 2-16-1 Tokiwadai, Ube, 755-8611, Japan It is not easy to detect character from comic because it is difficurt to recognise a picture drawn by human. We use Hough transform to detect figure even if figure s one copy is hidden because if you detect character from comic, you mistake detecting character or character s one copy is hidden by other thing. But, we have another problem that characters are drawn defferent little by little even if they are same size and aspect and comics have many things being not character. We propuse the method making new template by plural template. We suggest the method resolving difference of votes for radius by normalization. We improve the detection rate by detecting some frame and density of black because comic have several characteristic such as comic is devided image by frame, drawn by black and white and many expecting characters exist in comic. We show effectiveness by exparment using comic HUNTERXHUNTER volume 1, NARUTO volume 44 and Dr. SLUMP,volume 1. keyword: Hough Transform, Comic, Detecting Character, Patterm Matching 1
1 (3) (4) 3 3.1 1962 Hough [1] O Ballard r O [2] p r [3] R = r /r 1 R tmpr tmpr/r p r cx cy r rate = r /r θ x = r rate cos(θ)+cx 2 y = r rate sin(θ)+cy (1) x y 3.2 2 (1) 1 1 (2) 2
1: B Fig. 1 Length of black point from the reference point C, A = 1 C/B in template 1: 5 Tab. 1 Average detection rate of basic shape ( ) 4 87.9% 83.7% 59.5% 75.3% 6 4 22)/18.)+4 5 4.1 6 6 hunterxhunter number 1-8 5 45 FN.n3.".txt" HUNTERXHUNTER1 [4] NARUTO44 [5] 2 HUNTERXHUNTER 3 1 NARUTO 2: (HUNTERX- HUNTER) Fig. 2 Transition that number of votes for radius(hunterxhunter) 4.2 x f(x) f(x) =.1/25. exp(( x+ 4 35 3 25 2 15 5 FN.n1.".txt" FN.n8.".txt" 3
25 2 15 1 5 FN.n1.".txt" FN.n3.".txt" FN.n8.".txt" FN.n9.".txt" FN.n1.".txt" FN.n11.".txt" FN.n12.".txt" FN.n13.".txt" naruto number 1-13 5 5.1 3: (NARUTO) Fig. 3 Transition that number of votes for (2) radius(naruto) 1 8 6 4 2 p n aruto number 2-7 FN.n3.".txt" (1) 4: Fig. 4 Percentage of votes for radius 1./(1.+exp((-x/T1)/T 1 +x))+y 1 8 6 4 2 sigmoid1 x e14(x) 5: Fig. 5 Graph of approximating the percentage of vores 1.4 1.2 1.8.6.4.2 hunterxhunter number 1-8 FN.n3.".txt" 5.2 7 NARUTO 8 HUNTERX- HUNTER Dr.SLUMP1 [6] 6: Fig. 6 Normalization of votes for the radius of the difference 4
2 3: ( ) 3 4 Tab. 3 Distribution that orientation of face(right slanting face and other) HXH 14.3% 13.2% % % NARUTO 15.9% 14.8% 2.3% 1.1% Dr.SLUMP 16.3% 28.1% % % face(naruto) 7: (NARUTO) Fig. 7 Classification that orientation of 4: ( ) Tab. 4 Distribution that orientation of face(right upward slanting face and other) HXH % % % % NARUTO % 6.8% % 1.1% Dr.SLUMP.5% 2.% 1.5% % 8: (HUNTERXHUNTER) face(hunterxhunter) Fig. 8 Classification that orientation of 2: ( ) Tab. 2 Distribution that orientation of face(front face and other) HXH 36.3% 19.8% 16.5% % % NARUTO 43.2% 4.4% 9.1% % % Dr.SLUMP 14.3% 14.3% 22.7% % % 6 1 5.3 5
5 9 5: Tab. 5 Detection rate of frame in comic HXH NARUTO 53% 93% 6: Tab. 6 Detection rate of charcter HXH( ) 86% NARUTO( ) 86% Dr.SLUMP( ) 92% HXH( ) 7% NARUTO( ) 8% Dr.SLUMP( ) 8% 1 9: Fig. 9 Processing 6 1 7% 8% 7 [1] Hough, P. V. C. Method and maens for recognzing complex patterns, U. S. Patent No. 369654, 1962. [2] Ballard, D. H. Generalizing the Hough transform to detect arbitrary shapes, PR, Vol. 13, No. 2, pp. 111-122, 1981 [3] Stockman, G. C. and Agrawala, A. K. Equivalence of Hough Curve Detection to Template Matching, Com. of ACM, Vol. 2, No. 11, pp. 82-822, 1977 [4], HUNTERXHUNTER, 1,, 1998 [5], NARUTO, 44,, 28 [6], Dr.SLUMP, 1,, 198 1: Fig. 1 Result of detection 6