Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 Bsi Imge Fetures (BIFs) rising from pproximte Symmetry Type Lewis D. Griffin 1, M Lillholm 1, M Crosier 1, vn Snde (1) Computer Siene, University College London, London WC1E 6BT, UK. () Biomedil Engineering, Eindhoven University of Tehnology, the Netherlnds. l.griffin@s.ul..uk Abstrt. We onsider detetion of lol imge symmetry using liner filters. We prove simple riterion for determining if filter is sensitive to group of symmetries. We show tht derivtive-of-gussin (DtG) filters re exellent t deteting lol imge symmetry. Building on this, we propose very simple lgorithm tht, bsed on the responses of bnk of six DtG filters, lssifies eh lotion of n imge into one of seven Bsi Imge Fetures (BIFs). This effetively nd effiiently relizes Mrr s proposl for n imge priml sketh. We summrize results on the use of BIFs for texture lssifition, objet tegory detetion, nd pixel lssifition. Keywords: Gussin Derivtives; Hermite Trnsform; Group Theory. 1 Introdution Previous shemes for detetion of imge symmetry re firly omplex [1-6]; requiring, for exmple, omprison of the outputs of filters t multiple positions. Herein we show tht symmetries my be deteted by single liner filters. Building on this we present simple lgorithm tht omputes Mrr-type priml sketh [7] by tegorizing lol imge struture ording to its pproximte symmetry. The pper is orgnized s follows. In setion we present results on imge symmetries. In 3 we show how to test whether liner filter is sensitive to symmetry. In 4 we review imge mesurement with derivtive-of-gussin (DtG) filters. In 5 we onsider the symmetry-sensitivity of these DtG filters. In 6 we show how this sensitivity gives rise to system of Bsi Imge Fetures (BIFs). In 7 we summrize results on using BIFs for texture tegoriztion, objet tegory detetion nd pixel lssifition. In 8 we onlude. Setions -5 re distilltion of work published, in press nd under review in fuller form elsewhere [8-14]; 6 is new; prts of 7 hve been presented or re under review in fuller form elsewhere [9, 11].
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 Imge Symmetries Symmetry of struture (X) is lwys reltive to some lss of dmissible trnsformtions. A struture is sid to hve symmetry when non-trivil group of dmissible trnsformtions, known s the utomorphism group, eh leve it Aut X : t t X X. indistinguishble from the originl. This is denoted by Considering imges, n obvious lss of trnsformtions re the sptil isometries; nd the possible symmetries, reltive to this lss, hve long been tlogued [15-17]. A broder lss of trnsformtions, where eh sptil isometry is ombined with permuttion of finite set of imge olour vlues, hs lso been onsidered. These llow the symmetries of, for exmple, Esher s Reptiles to be expressed [18]. The gmut of possible olour symmetries hs been fully determined [19, 0]. We propose tht the lss of imge isometries, defined s sptil isometry ombined with n intensity isometry, is pproprite for imges. We write n imge g i, s, where i : is n intensity isometry, nd s : is isometry s sptil isometry. Suh n imge isometry is pplied to n imge ording to g I i Is _ i Is_ I :. Choosing lss of trnsformtions is tntmount to hoosing geometry [1], nd the geometry tht orresponds to the lss of imge isometries hs previously been onsidered for imges [] nd muh erlier, bstrtly, s one of lrger lss of possible geometries [3]. We hve employed method for determining the possible utomorphism groups of imges, reltive to the lss of imge isometries. The method relies on two results. First, tht the projetion of group of imge isometries onto their sptil or intensity omponents in both ses mkes group. Seond, tht (exept for speil se) the intensity projetion group must be isomorphi to ftor group of the sptil projetion group [8]. Using the method, we hve determined the possible utomorphism groups of -D imges, exept for ses tht ontin disrete periodi trnsltions. A summry of these possible symmetries, together with our nottionl system is shown in fig. 1. The symmetries inlude: fmilir ones, suh s refletionl (,1 ), reflet-nd-negte ( 6,1 ), nd Yin-Yng type ( 7, ); simple but often ignored ones, suh s vrition in one diretion only ( 3 ); simple but novel, suh s ontinuous trnslte-nd-inrement in one diretion, plus line of refletion prllel to tht diretion ( 11 ); nd some wholly novel, suh s ontinuous trnslte-nd-inrement in one diretion, plus ontinuous line of entres of Yin-Yng type symmetry ( 1 ). 3 Sensitivity of Liner Filters to Symmetries Detetion of symmetry seems to require multiple mesurements, but this is inorret. Consider +1/-1 filter, suh s used in finite-differene shemes. When positioned so tht it strddles puttive line of refletion, neessry riterion for the symmetry is tht the filter gives 0 response. We generlizes this: filter F is sensitive to symmetry K if it gives the sme response to ll imges tht hve the
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 symmetry (i.e. f AutI K F I f ). This definition is imprtil beuse it requires ssessment ross ll imges. However, we hve found neessry nd suffiient test tht requires only single integrl to be omputed. We present this below in Theorem 1, fter introduing some nottion.
6,3, 6,5, Out[81]= 1,3, 1,5, Out[90]= 6,4, 6,6, Out[8]= 1,4, 1,6, Out[85]= Out[80]=,4,,6,,3,,5, Out[9]= 8,4, 8,6, 7,4, 7,6, Fig. 1. The group/subgroup lttie of the possible imge symmetries, exluding those with disrete periodi trnsltion. Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 Out[93]= Out[94]= Out[96]= 6, 6,1 Out[89]= Out[86]= 5 Out[84]= 1, Out[97]= Out[83]= Out[79]= Out[99]= onst Out[9]= 4 9, Out[95]= 8, Out[101]= Out[87]= Out[91]=,1 3 7, 0 Out[103]= 11 Out[100]= slope Out[10]= 10 1
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 We use n inner produt nottion ( F I : F x I x ) to denote the mesurement of n imge opertor F K i, sk x I : by filter F : ; nd we define n : i s F whih, roughly speking, smers filter by group. Theorem - Symmetry-Sensitivity Test for Filters F is sensitive to K if nd only if K FF 0 Proof A forml proof will be published elsewhere [14]. Intuitively the truth of the theorem n be understood s follows. The signl tht filter sees best is opy of itself. Of ll the symmetri signls, symmetrised version of the filter should be the most esily seen. If the filter nnot see symmetrised version of itself, then it is insensitive to the symmetry. 4 Gussin Derivtive Filters Gussin Derivtive (DtG) filters re defined in 1-D by x n n d 1 x n n 1 n :, : G x e G x G x H G x dx where H n is the nth Hermite polynomil; nd in -D by m n m n, :, G x y G x G y. They re used s generl-purpose method to probe n imge lotion (whih for simpliity we ssume is t the origin 0 ) by omputtion of inner produts mn mn, j G I. Typilly, one mesures with fmily of DtG filters up to some order e.g. the nd order fmily mn, G 0 m n. Sle-normlized filter responses pq : p q jpq mke lter equtions simpler. The suitbility of DtGs s the front-end of n unommitted omputtionl vision system rises from the symmetries tht individul filters nd fmilies possess [4]. First mongst these is sle symmetry, whih mnifests s hnge of size, but not of shpe, when DtG is resled by blurring with Gussin kernel. Seond is tht the liner spn of fmily of DtGs is rottionlly symmetri. The responses of bnk of DtG filters entngle intrinsi nd extrinsi spets of imge struture. For exmple, n in-plne rottion of the imge bout the mesurement point uses the DtG responses to hnge. A representtion tht disentngles these spets for mesurement up to nd order hs been developed [13]. The representtion works by ftoring out of the 6-D nd order DtG response spe the hnges due to the group of imge isometries tht fix the mesurement point nd do
Out[81]= Out[80]= Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 not invert the intensity xis whih we denote D A 1 0. 1.1 = -0.1 -.3 0.8 4.1-3.7 imge pth nd order DtG fmily point in 6-D jet spe 6-D jet spe D A 0 1 - The group of entred rottions nd refletions, nd positive ffine intensity re-slings = The nd order lol-imge-struture orbifold Fig.. The top prt illustrtes shemtilly the probing of n imge pth by bnk of DtG filters resulting in point in jet spe; the bottom, the ftoring of the jet spe by group of trnsformtions resulting in the lol-imge-struture orbifold. The result is n orbifold type of mnifold with boundries, reses nd orners llowed onsisting of 3-D nd 0-D omponent (figure ). The intrinsi spet of 6-tuple of filter responses orresponds to prtiulr lotion in the orbifold, nd is invrint to rotting the imge bout the mesurement point, refleting it in line through the mesurement point, or ffinely sling the intensity. When the responses of the 1 st nd nd order DtG filters re ll zero, the intrinsi spet is the 0-D prt of the orbifold; ll other responses mp to the 3-D omponent. A oordinte system l, b,, 0, 0, for the 3-D omponent is given by [13]: l rtn 4 4, 10 01 0 0 11 0 0 b rtn, 4 1 10 01 0 0 11 01 10 0 0 10 01 11 01 10 11 10 01 0 0 rtn 4, The orbifold hs been equipped with metri, indued by one on the filter response spe, whih expressed s line element in the lb-system is 1 ds dl os l db d 5 3osb sin b. The orbifold is intrinsilly urved, but it n be embedded into Euliden 3-spe with only mild distortion.
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 5 Symmetry-sensitivity of DtG Filters Using the elements of setions -4, we n determine whih DtG filters re sensitive to whih symmetries. We onsider not just nonil filter forms (e.g. n x-derivtive) but ny liner ombintion of filters in the nd order filter fmily. This llows us to determine the symmetry-sensitivity of the entire filter fmily, independent of the prtiulr bsis filters used. For exmple, while the x-derivtive filter is sensitive to refletionl symmetry with vertil mirror line through the mesurement point, the x- nd y-derivtives together re sensitive to ny refletionl symmetry in line through the mesurement point. 6, 3, 7, 6, 8,4 6, 6, 7, 1, 8, 6, onst 1, 4, 1, 3,,3 5 6, 8,4 4 7, 7,4 8, 9 1 slope, 1, 11 3 g 4 g 9,1, 8, 5 g 11 10 g 1 6,1 6, 8, SS is the exterior only 3 0 SS is the entire volume g g 1,,1 g g 6,1 7, Fig 3. The sensitivity-submnifolds (SS) of different symmetry types re shown in red. The different possible SS re rrnged in lttie indued by inlusion reltions. The symmetry type lbels orrespond to those used in figure 1. Supersripts indite the sptil reltionship between the symmetry nd the origin: indites origin-entred rottion; n + tht the origin is ontined in line of refletion, but is not entre of rottion; similrly for - nd nti-refletions; g indites generl position, neither entred nor ligned. All symmetries lbelled in box hve the indited SS; those on the left re miniml. g 8, The filter fmily sensitivities n be projeted into the orbifold to determine where the intrinsi omponent of the jet responses must lie whenever the imge hs ny of
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 lss of symmetries equivlent by onjugtion with n element of D A 1 0. We ll the restrited set of possible responses the sensitivity-submnifold (SS). For exmple the SS is the orbifold exterior ( 0 ) for refletionl symmetry in mirror through the mesurement point. The results re summrized in fig 3. 6 Symmetry-bsed Bsi Imge Fetures (BIFs) We hve used the symmetry sensitivities of the DtG filters s strting point in defining set of Bsi Imge Fetures (BIFs) tht relize Mrr s ide of priml sketh of imge struture, in omputtionlly simple sheme. We do not lim tht the sheme is derived s rigorously s the results on symmetry sensitivity. Our sheme works by onsidering the orbifold projetion of jets, nd lssifying them ording to the SS tht they re losest to i.e. we define Voronoi ell prtitioning of the orbifold with the SS s ell entres. We find tht this works best when only seven 0-D SS (the first nd seond rows of figure 3) re used, though we nnot justify this beyond tht it produes nie results. The resulting orbifold prtitioning is shown in the top-left of figure 4. Fig. 4. Top left: the prtitioning of the orbifold into BIF tegories. Bottom left: BIFs lulted ross rnge of sles for simple imge of figure 8 ; in eh ube sle inreses right-to-left. Lower ubes setioned for visulistion. Right: n exmple omplex greysle imge, with BIFs lulted t one prtiulr sle.
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 The orbifold distne to the six of these SS tht lie in the 3-D omponent of the orbifold re simple to ompute; for exmple, the distne to the 7, SS is 1 1 0 11 0 10 01 1 tn. To find whih distne is shortest it is omputtionlly equivlent but simpler to find whih of six quntities is mximum. The distne to the seventh SS, whih orresponds to the origin of jet spe where ll the 1 st nd nd derivtive filters hve zero response, is not well-defined. We inorporte it into our sheme by using multiple of the 0 th order jet response. The full resulting sheme for omputing BIFs is s follows. i) ompute sle-normlized DtG filter responses s desribed in setion 7. 1 1 ii) ompute : nd : 0 0 4 0 0 11 iii) lssify ording to whih is the lrgest of 00 10 01 M.,,,,,,. In our sheme the only free prmeters, tht hve to be tuned to the pplition re the filter sle nd whih ontrols the mount of imge lssified s flt; setting of 0.05 is n effetive defult. For disply purposes we find the following olour sheme effetive: if.00 is the lrgest of M then olour the pixel pink; if 10 01 is lrgest olour it grey; then blk, white, blue, yellow nd green. 7 Exmple pplitions using BIFs We summrize results on using BIFs for texture, objet nd pixel lssifition. 7.1 Texture lssifition Textures re often lssified bsed on representtion of them by histogrm over texton vobulry [5-9]. Textons re tegoril pth lssifitions [5, 30]. To define the texton vobulry, spe of pth desriptions is typilly Voronoi prtitioned into on-the-order-of 1000 texton tegories, usully round entres found by k-mens lustering of the responses from mny imges. Textures re then lssified by nerest-neighbour mthing of histogrms. We hve investigted the lssifition performne of n pproh in whih imges re lbelled using sptil omplexes of BIFs insted of Voronoi ells in lol desription vetor spe. Our pproh is (i) simpler beuse we hve eliminted the lustering step needed to produe ditionry of fetures, nd (ii) fster beuse we ssign imge pthes to histogrm bins without hving to use high-dimensionl nerest-neighbour omputtion. We ll the sptil omplexes of BIFs tht we use nlogously to textons, Bsi Imge Ptterns (BIPs). The type of BIP tht we hve found effetive for texture
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 desription is sle-templte of the BIFs t the sme lotion but t four, otve-seprted sles. Unlike sptil-templte BIPs, these sle-templtes retin the rottion invrine of BIFs, whih hs been shown [30] to be dvntgeous in texture lssifition tsks. For textures, we do not use the pink/flt BIF tegory, so four sles produes 6 4 =196 bin histogrm representtion, whih seems to pture the right trde-off between speifiity nd generlity (see figure 5). 0.05 0.04 0.03 0.0 0.01 11 13 14 1345678910 15 16 17 18 19 0 1 4 5 6 7 8 9 30 31 3 33 35 36 37 38 39 40 41 4 43 44 45 46 47 48 49 50 51 5 53 54 55 56 57 58 59 60 61 6 63 64 65 66 67 68 69 70 71 7 73 74 75 76 77 78 79 80 81 8 83 84 85 86 87 100 101 10 103 104 105 106 88 89 90 91 9 93 94 95 107 108 109 110 111 11 113 96 97 98 99 114 115 116 117 118 119 10 11 1 13 14 15 16 17 18 19 130 131 13 133 134 135 136 137 138 139 140 141 14 143 144 145 146 147 148 149 150 151 15 153 154 155 156 157 158 159 160 161 16 163 164 165 166 167 168 169 170 171 17 173 174 175 176 177 178 179 180 181 18 183 184 185 186 187 188 189 190 191 19 193 194 195 196 197 198 199 00 01 0 03 04 05 06 07 08 09 10 11 1 13 14 15 16 17 18 19 0 1 3 4 5 6 7 8 9 30 31 3 33 34 35 36 37 38 39 40 41 4 43 44 45 46 47 48 49 50 51 5 53 54 55 56 57 58 59 60 61 6 63 64 65 66 67 68 69 70 71 7 73 74 75 76 77 78 79 80 81 8 83 84 85 86 87 88 89 90 91 9 93 94 95 96 97 98 99 300 301 30 303 304 305 306 307 308 309 310 311 31 313 314 315 316 317 318 319 30 31 3 33 34 35 36 37 38 39 330 331 33 333 334 335 336 337 338 339 340 341 34 343 344 345 346 347 348 349 350 351 35 353 354 355 356 357 358 359 360 361 36 363 364 365 366 367 368 369 370 371 37 373 374 375 376 377 378 379 380 381 38 383 384 385 386 387 388 389 390 391 39 393 394 395 396 397 398 399 400 401 40 403 404 405 406 407 408 409 410 411 41 413 414 415 416 417 418 419 40 41 4 43 44 45 46 47 48 49 430 431 43 433 434 435 436 437 438 439 440 441 44 443 444 445 446 447 448 449 450 451 45 453 454 455 456 457 458 459 460 461 46 463 464 465 466 467 468 469 470 471 47 473 474 475 476 477 478 479 480 481 48 483 484 485 486 487 488 489 490 491 49 493 494 495 496 497 498 499 500 501 50 503 504 505 506 507 508 509 510 511 51 513 514 515 516 517 518 519 50 51 5 53 54 55 56 57 58 59 530 531 53 533 534 535 536 537 538 539 540 541 54 543 544 545 546 547 548 549 550 551 55 553 554 555 556 557 558 559 560 561 56 563 564 565 566 567 568 569 570 571 57 573 574 575 576 577 578 579 580 581 58 583 584 585 586 587 588 589 590 591 59 593 594 595 596 597 598 599 600 601 60 603 604 605 606 607 608 609 610 611 61 613 614 615 616 617 618 619 60 61 6 63 64 65 66 67 68 69 630 631 63 633 634 635 636 637 638 639 640 641 64 643 644 645 646 647 648 649 650 651 65 653 654 655 656 657 658 659 660 661 66 663 664 665 666 667 668 669 670 671 67 673 674 675 676 677 678 679 680 681 68 683 684 685 686 687 688 689 690 691 69 693 694 695 696 697 698 699 700 701 70 703 704 705 706 707 708 709 710 711 71 713 714 715 716 717 718 719 70 71 7 73 74 75 76 77 78 79 730 731 73 733 734 735 736 737 738 739 740 741 74 743 744 745 746 747 748 749 750 751 75 753 754 755 756 757 758 759 760 761 76 763 764 765 766 767 768 769 770 771 77 773 774 775 776 777 778 779 780 781 78 783 784 785 786 787 788 789 790 791 79 793 794 795 796 797 798 799 800 801 80 803 804 805 806 807 808 809 810 811 81 813 814 815 816 817 818 819 80 81 8 83 84 85 86 87 88 89 830 831 83 833 834 835 836 837 838 839 840 841 84 843 844 845 846 847 848 849 850 851 85 853 854 855 856 857 858 859 860 861 86 863 864 865 866 867 868 869 870 871 87 873 874 875 876 877 878 879 880 881 88 883 884 885 886 887 888 889 890 891 89 893 894 895 896 897 898 899 900 901 90 903 904 905 906 907 908 909 910 911 91 913 914 915 916 917 918 919 90 91 9 93 94 95 96 97 98 99 930 931 93 933 934 935 936 937 938 939 940 941 94 943 944 945 946 947 948 949 950 951 95 953 954 955 956 957 958 959 960 961 96 963 964 965 966 967 968 969 970 971 97 973 974 975 976 977 978 979 980 981 98 983 984 1000 1001 100 985 986 987 988 989 990 991 1003 1004 1005 1006 1007 1008 1009 1010 99 993 994 995 996 997 998 999 1011 101 1013 1014 1015 1016 1017 1018 1019 100 101 10 103 104 105 106 107 108 109 1030 1031 103 1033 1034 1035 1036 1037 1038 1039 1040 1041 104 1043 1044 1045 1046 1047 1048 1049 1050 1051 105 1053 1054 1055 1056 1057 1058 1059 1060 1061 106 1063 1064 1065 1066 1067 1068 1069 1070 1071 107 1073 1074 1075 1076 1077 1078 1079 1080 1081 108 1083 1084 1085 1086 1087 1088 1089 1090 1091 109 1093 1094 1095 1096 1097 1098 1099 1100 1101 110 1103 1104 1105 1106 1107 1108 1109 1110 1111 111 1113 1114 1115 1116 1117 1118 1119 110 111 11 113 114 115 116 117 118 119 1130 1131 113 1133 1134 1135 1136 1137 1138 1139 1140 1141 114 1143 1144 1145 1146 1147 1148 1149 1150 1151 115 1153 1154 1155 1156 1157 1158 1159 1160 1161 116 1163 1164 1165 1166 1167 1168 1169 1170 1171 117 1173 1174 1175 1176 1177 1178 1179 1180 1181 118 1183 1184 1185 1186 1187 1188 1189 1190 1191 119 1193 1194 1195 1196 1197 1198 1199 100 101 10 103 104 105 106 107 108 109 110 111 11 113 114 115 116 117 118 119 10 11 1 13 14 15 16 17 18 19 130 131 13 133 134 135 136 137 138 139 140 141 14 143 144 145 146 147 148 149 150 151 15 153 154 155 156 157 158 159 160 161 16 163 164 165 166 167 168 169 170 171 17 173 174 175 176 177 178 179 180 181 18 183 184 185 186 187 188 189 190 191 19 193 194 195 196 197 198 199 1300 1301 130 1303 1304 1305 1306 1307 1308 1309 1310 1311 131 1313 1314 1315 1316 1317 1318 1319 130 131 13 133 134 135 136 137 138 139 1330 1331 133 1333 1334 1335 1336 1337 1338 1339 1340 1341 134 1343 1344 1345 1346 1347 1348 1349 1350 1351 135 1353 1354 1355 1356 1357 1358 1359 1360 1361 136 1363 1364 1365 1366 1367 1368 1369 1370 1371 137 1373 1374 1375 1376 1377 1378 1379 1380 1381 138 1383 1384 1385 1386 1387 1388 1389 1390 1391 139 1393 1394 1395 1396 1397 1398 1399 1400 1401 140 1403 1404 1405 1406 1407 1408 1409 1410 1411 141 1413 1414 1415 1416 1417 1418 1419 140 141 14 143 144 145 146 147 148 149 1430 1431 143 1433 1434 1435 1436 1437 1438 1439 1440 1441 144 1443 1444 1445 1446 1447 1448 1449 1450 1451 145 1453 1454 1455 1456 1457 1458 1459 1460 1461 146 1463 1464 1465 1466 1467 1468 1469 1470 1471 147 1473 1474 1475 1476 1477 1478 1479 1480 1481 148 1483 1484 1485 1486 1487 1488 1489 1490 1491 149 1493 1494 1495 1496 1497 1498 1499 1500 1501 150 1503 1504 1505 1506 1507 1508 1509 1510 1511 151 1513 1514 1515 1516 1517 1518 1519 150 151 15 153 154 155 156 157 158 159 1530 1531 153 1533 1534 1535 1536 1537 1538 1539 1540 1541 154 1543 1544 1545 1546 1547 1548 1549 1550 1551 155 1553 1554 1555 1556 1557 1558 1559 1560 1561 156 1563 1564 1565 1566 1567 1568 1569 1570 1571 157 1573 1574 1575 1576 1577 1578 1579 1580 1581 158 1583 1584 1585 1586 1587 1588 1589 1590 1591 159 1593 1594 1595 1596 1597 1598 1599 1600 1601 160 1603 1604 1605 1606 1607 1608 1609 1610 1611 161 1613 1614 1615 1616 1617 1618 1619 160 161 16 163 164 165 166 167 168 169 1630 1631 163 1633 1634 1635 1636 1637 1638 1639 1640 1641 164 1643 1644 1645 1646 1647 1648 1649 1650 1651 165 1653 1654 1655 1656 1657 1658 1659 1660 1661 166 1663 1664 1665 1666 1667 1668 1669 1670 1671 167 1673 1674 1675 1676 1677 1678 1679 1680 1681 168 1683 1684 1685 1686 1687 1688 1689 1690 1691 169 1693 1694 1695 1696 1697 1698 1699 1700 1701 170 1703 1704 1705 1706 1707 1708 1709 1710 1711 171 1713 1714 1715 1716 1717 1718 1719 170 171 17 173 174 175 176 177 178 179 1730 1731 173 1733 1734 1735 1736 1737 1738 1739 1740 1741 174 1743 1744 1745 1746 1747 1748 1749 1750 1751 175 1753 1754 1755 1756 1757 1758 1759 1760 1761 176 1763 1764 1765 1766 1767 1768 1769 1770 1771 177 1773 1774 1775 1776 1777 1778 1779 1780 1781 178 1783 1784 1785 1786 1787 1788 1789 1790 1791 179 1793 1794 1795 1796 1797 1798 1799 1800 1801 180 1803 1804 1805 1806 1807 1808 1809 1810 1811 181 1813 1814 1815 1816 1817 1818 1819 180 181 18 183 184 185 186 187 188 189 1830 1831 183 1833 1834 1835 1836 1837 1838 1839 1840 1841 184 1843 1844 1845 1846 1847 1848 1849 1850 1851 185 1853 1854 1855 1856 1857 1858 1859 1860 1861 186 1863 1864 1865 1866 1867 1868 1869 1870 1871 187 1873 1874 1875 1876 1877 1878 1879 1880 1881 188 1883 1884 1885 1886 1887 1888 1889 1890 1891 189 1893 1894 1895 1896 1897 1898 1899 1900 1901 190 1903 1904 1905 1906 1907 1908 1909 1910 1911 191 1913 1914 1915 1916 1917 1918 1919 190 191 19 193 194 195 196 197 198 199 1930 1931 193 1933 1934 1935 1936 1937 1938 1939 1940 1941 194 1943 1944 1945 1946 1947 1948 1949 1950 1951 195 1953 1954 1955 1956 1957 1958 1959 1960 1961 196 1963 1964 1965 1966 1967 1968 1969 1970 1971 197 1973 1974 1975 1976 1977 1978 1979 1980 1981 198 1983 1984 1985 1986 1987 1988 1989 1990 1991 199 1993 1994 1995 1996 1997 1998 1999 000 001 00 003 004 005 006 007 008 009 010 011 01 013 014 015 016 017 018 019 00 01 0 03 04 05 06 07 08 09 030 031 03 033 034 035 036 037 038 039 040 041 04 043 044 045 046 047 048 049 050 051 05 053 054 055 056 057 058 059 060 061 06 063 064 065 066 067 068 069 070 071 07 073 074 075 076 077 078 079 080 081 08 083 084 085 086 087 088 089 090 091 09 093 094 095 096 097 098 099 100 101 10 103 104 105 106 107 108 109 110 111 11 113 114 115 116 117 118 119 10 11 1 13 14 15 16 17 18 19 130 131 13 133 134 135 136 137 138 139 140 141 14 143 144 145 146 147 148 149 150 151 15 153 154 155 156 157 158 159 160 161 16 163 164 165 166 167 168 169 170 171 17 173 174 175 176 177 178 179 180 181 18 183 184 185 186 187 188 189 190 191 19 193 194 195 196 197 198 199 00 01 0 03 04 05 06 07 08 09 10 11 1 13 14 15 16 17 18 19 0 1 3 4 5 6 7 8 9 30 31 3 33 34 35 36 37 38 39 40 41 4 43 44 45 46 47 48 49 50 51 5 53 54 55 56 57 58 59 60 61 6 63 64 65 66 67 68 69 70 71 7 73 74 75 76 77 78 79 80 81 8 83 84 85 86 87 88 89 90 91 9 93 94 95 96 97 98 99 300 301 30 303 304 305 306 307 308 309 310 311 31 313 314 315 316 317 318 319 30 31 3 33 34 35 36 37 38 39 330 331 33 333 334 335 336 337 338 339 340 341 34 343 344 345 346 347 348 349 350 351 35 353 354 355 356 357 358 359 360 361 36 363 364 365 366 367 368 369 370 371 37 373 374 375 376 377 378 379 380 381 38 383 384 385 386 387 388 389 390 391 39 393 394 395 396 397 398 399 400 401 Fig. 5. Left: An imge from the CUReT 'polyester' texture lss. Centre: BIFs omputed t four otve-seprted sles, stked to form n rry of 'olumn-bips'. Right: Ourrene histogrm of olumn-bips from every position in the imge" Our method hs been tested on the CUReT texture dtset [31]. As reported in [9], the simple olumn-bip representtion nd nerest-neighbour mthing using the Bhtthryy distne orretly lssifies 98.±0.1% of the remining 49 imges per lss, whih is t lest s good s other methods using nerest-neighbour lssifiers. Extending this method by using multi-sle histogrm omprison [9] results in n improvement to 98.6±0.1% on CUReT, whih is omprble to methods [7] using SVMs for lssifition; nd produes wht re, to the best of our knowledge, the best reported results [9] on the more hllenging UIUCTex [3] nd KTH-TIPS [7] dtsets, whih inlude vritions in sle. 7. Objet Ctegoriztion Texton pprohes hve lso been shown to be useful for objet tegoriztion [8, 33]. Similr to texture, the stndrd pproh is to prtition pth desriptor spe, suh s tht used by SIFT [34, 35], into on-the-order-of 1000 tegories (visul words) nd then to desribe eh imge to be nlyzed by wht visul words it ontins, nd to use mhine lerning tehniques to determine lssifier tht n predit the tegory of objet bsed on suh desriptions. We hve onduted preliminry experiments to ssess whether visul words built from BIFs ould be used rther thn SIFT-spe tegories. As with texture this would be simpler nd fster. For our initil experiments, we hve lbelled pixels ording to their BIF type nd, inspired by SIFT, with n orienttion, quntized t the 4 level. The orienttion depends on the BIF type: grey BIFs hve one of eight possible orienttions bsed on 1 st order struture; yellow, green nd blue BIFs hve
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 one of four possible orienttions bsed on nd order struture; blk, white nd pink BIFs re unoriented. Thus we hve twenty-three possible orienttion-ugmented BIF (obif) lbels. obifs re nturl nd order generliztion of the grdient orienttion lphbet typilly used in SIFT [34, 35]. See figure 6 for n exmple imge nd lulted obifs. Out[70]= Out[17]= Fig. 6. The top row (left) shows n imge from the PASCAL hllenge, lbeled with diretion-ugmented BIFs t right. On the bottom re shown the 44 templte BIPs whose ourrene in n imge most informtively signls the presene of r. We hve tested three different types of visul word, whih when built from BIFs or obifs we ll Bsi Imge Ptterns (BIPs); two bsed on geometril prtitioning of pth spe nd one bsed on more stndrd dt-driven quntiztion. Eh BIP system hs been used with simple un-optimised of-the-shelf lssifiers nd pplied to the 0-lss PASCAL VOC 008 [33] objet reognition hllenge dtset. Our sore in figure 7 is bsed on lte fusion of the three shemes nd is mid-field: bove other first-time entrnts nd below well-optimised vetern entries. Using the PASCAL VOC 008 [33] dtset, exmples of 4x4 templte BIPs whose presene in imges is pproximtely independent, nd whih re mximlly informtive for the r tegory re shown in figure 6.
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 SurreyUv A SRKDA Uv A TreeSFS LEAR shotgun Uv A FullSFS Uv A Soft5ColorSift LEAR flt XRCE TKK ALL SFBS TKK MAXVAL BerlinFIRSTNikon UCL ECPLIAMA CASIA LinSVM INRIASly CMA CASIA NonLinSVM INRIASly MEVO FIRST SCST FIRST SC1C CASIA NeurlNet 0 10 0 30 40 50 Fig. 7. Results for the PASCAL VOC 008 hllenge. Eh br in the hrt is hllenge entry - our result is highlighted. 7.3 Pixel Clssifition Mny imge problems involve inferring one of smll lss of lbels for eh pixel of n imge. For omplex imges with unpreditble globl struture, most pprohes blne the likelihood of the lbels, given the lol imge struture, nd the likelihood of the lol rrngement of inferred lbels. In both ses the likelihoods re omputed on the bsis of sttistis lernt from groundtruth-lbelled trining dt. We hve experimented with the use of BIFs in the omputtion of lbel likelihoods given the imge i.e. ignoring the likelihood of rrngements of inferred lbels. For our experiments we hve used -D Eletron Mirosopy imges of neuronl grey mtter tissue, stined to enhne neuronl membrnes. We trined on four imges with hnd-drwn groundtruth dt, inditing the position of membrnes, nd evluted on further four imges. We use k-nerest Neighbour (k-nn) pproh to lssifition. NN lssifition strts by ompiling list of desriptors of ll the pthes in the trining dt, together with the groundtruth lbel of the pixel in the pth entre. The lssifier is used by extrting pth round eh pixel to be lssified, forming desription of it, ompring the desription to eh the ompiled desriptions, finding the k whih re most similr, nd ssigning the pixel being nlyzed with the lbel ssoited with the mjority of the k. We evluted bseline solution bsed on pixel vlues. The distne between two pth desriptions is simply the Euliden distne between their blurred pixel vlues, minimized over llowing one pth to be rotted or refleted into eight onfigurtions. We jointly optimize blur, pth size nd k. The best settings tht we find re: no blur, 77 pthes, nd k=14. At these settings membrne-lbelled pixels overlp (intersetion divided by union) with the groundtruth by 48%. Our solution uses pth of BIF lbels s pth desriptor. The distne between two desriptors is simply the number of pixels where the lbel does not gree. As in
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 the bseline, we minimize the distne over one of the pthes being rotted or refleted. We jointly optimize the sle ( ) t whih the BIFs re omputed, the prmeter whih ontrols the mount of the flt BIF lss, pth size nd k. The best settings tht we find re 1., 0.15, 99 pthes, nd k=10. At these settings we hieve n overlp of 55%. See figure 8. imge groundtruth greylevel-bsed lssifition BIF-bsed lssifition BIFs Fig. 8. Typil results of our pixel-lssifition system. So, using BIF- rther thn greylevel-desription rises the sore from 48% to 55%. Computtion is lso fster beuse the knn lookup domintes the ost of omputing pth desriptions, nd with BIFs the distnes tht need to be omputed re of Hmming rther thn Euliden type. 8 Conlusions We hve derived sheme for lssifying imge struture into one of seven BIF types bsed on the outputs of bnk of six DtG filters. Applied to n entire imge, the output relizes Mrr s notion of n imge priml sketh. Presented results show tht BIF desription is simple, fst nd effetive for texture, objet nd pixel lssifition. The BIF system ws derived by onsidering the sensitivity of DtG filters to imge symmetry. Although the finl lgorithm is plesingly simple, there re some wek points in the derivtion of the BIFs from symmetry sensitivities. Speifilly, why re only the 0-D SS onsidered, how extly does orbifold distne orrespond to degree of filure of symmetry, why should lest-pproximte lol symmetry be n effetive feture lbel? We hope tht the foundtion of symmetry-sensitivity of DtGs n eventully nswer ll of these questions in sheme where rbitrry hoie hs been eliminted. Suh sheme will be extendble to higher-order filter fmilies (where ppel to visul evidene nd pst prtie re less effetive), for whih riher lphbet of feture lbels is to be expeted. We predit tht suh riher lphbet will give more effetive solutions in the pplition res tht we hve reviewed. Referenes 1. Liu, Y.X., R.T. Collins, nd Y.H. Tsin, A omputtionl model for periodi pttern pereption bsed on frieze nd wllpper groups. IEEE Trnstions on
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 Pttern Anlysis nd Mhine Intelligene, 004. 6(3): p. 354-371.. Sognmillo, R., et l., A feture-bsed model of symmetry detetion. Proeedings of the Royl Soiety of London Series B-Biologil Sienes, 003. 70(155): p. 177-1733. 3. Mellor, M. nd M. Brdy, A new tehnique for lol symmetry estimtion, in Pro. Sle Spe & PDE Methods in Comp Vis. LNCS vol. 3459. 005. p. 38-49. 4. Bonneh, Y., D. Reisfeld, nd Y. Yeshurun, Quntifition of lol symmetry - pplition to texture-disrimintion. Sptil Vision, 1994. 8(4): p. 515-530. 5. Mnini, S., S.L. Slly, nd R. Gurnsey, Detetion of symmetry nd ntisymmetry. Vision Reserh, 005. 45(16): p. 145-160. 6. Bylis, G.C. nd. Driver, Pereption of symmetry nd repetition within nd ross visul shpes: Prt-desriptions nd objet-bsed ttention. Visul Cognition, 001. 8(): p. 163-196. 7. Mrr, D., Vision. 198, New York: W H Freemn & o. 8. Griffin, L.D., Symmetries of 1-D Imges. ournl of Mthemtil Imging nd Vision, 008. 31(-3): p. 157-164. 9. Crosier, M. nd L.D. Griffin, Texture lssifition with ditionry of bsi imge fetures, in CVPR '08. 008, IEEE. 10. Lillholm, M. nd L.D. Griffin, Sttistis nd tegory systems for the shpe index desriptor of lol imge. Imge nd Vision Computing, 008. in press. 11. Lillholm, M. nd L.D. Griffin, Novel imge feture lphbets for objet reognition, in ICPR '08. 008. 1. Griffin, L.D., Symmetries of D imges: ses without periodi trnsltion. ournl of Mthemtil Imging nd Vision, in press. 13. Griffin, L.D., The nd order lol-imge-struture solid. IEEE Trnstions on Pttern Anlysis nd Mhine Intelligene, 007. 9(8): p. 1355-1366. 14. Griffin, L.D. nd M. Lillholm, Symmetry-sensitivity of derivtive of gussin filters. IEEE Trnstions on Pttern Anlysis nd Mhine Intelligene, in press. 15. Bieberbh, L., Über die bewegungsgruppen der euklidishen rume I. Mthemtishe Annlen, 1911. 70: p. 97. 16. Conwy,.H., et l., On three-dimensionl spe groups. Contributions to Algebr nd Geometry, 001. 4(): p. 475-507. 17. Grünbum, B. nd G.C. Shephrd, Tilings nd Ptterns. 1987, New York: WH Freemn & o. 18. Shttshneider, D., MC Esher. Visions of Symmetry. 1990: Plenum Press. 19. Holser, W.T., Clssifition of symmetry groups. At Crystllogrphi, 1961. 14: p. 136-14. 0. Loeb, A.A., Color nd Symmetry. 1978: Robert E. Krieger. 1. Klein, F., A omprtive review of reent reserhes in geometry (trns. by MW Hskell). Bulletin of the New York Mthemtil Soiety, 189. : p. 15-49.. Koenderink,.. nd A.. vn Doorn. Imge proessing done right. in ECCV 00. 00. Copenhgen: Springer. 3. Cyley, A., Sixth memoir upon the quntis. Philosophil Trnstions of the Royl Soiety, 1859. 149: p. 61-70. 4. Koenderink,.. nd A.. vn Doorn, Generi Neighborhood Opertors. Ieee Trnstions on Pttern Anlysis nd Mhine Intelligene, 199. 14(6): p. 597-
Author s Version: Griffin LD, Lillholm M, Crosier M & vn Snde (009) Bsi Imge Fetures (BIFs) rising from lol symmetry type. In: Pro. SSVM 09. LNCS vol. 5567:343-355 605. 5. Vrm, M. nd A. Zissermn, Texture lssifition: re filter bnks neessry?, in CVPR '03. 003, IEEE. 6. Vrm, M. nd A. Zissermn, A sttistil pproh to texture lssifition from single imges. Interntionl ournl of Computer Vision, 005. 6(1): p. 61-81. 7. Hymn, E., et l., On the signifigne of rel-world onditions for mteril lssifition, in ECCV '04. 004, Springer. p. 53-66. 8. Zhng,., et l., Lol fetures nd kernels for lssifition of texture nd objet tegories: omprehensive study, in CVPR '06. 006. 9. Perronnin, F., et l., Adpted vobulries for generi visul tegoriztion, in ECCV '06. 006. p. 464-475. 30. Vrm, M. nd A. Zissermn, Unifying Sttistil Texture Clssifition Frmeworks. Imge nd Vision Computing, 005. In Press. 31. Cul, O.G. nd K.. Dn, Compt representtion of bidiretionl texture funtions, in CVPR '01. 001, IEEE. 3. Lzebnik, S.C., C. Shmid, nd. Pone, A spre texture representtion using lol ffine regions. IEEE Trnstions on Pttern Anlysis nd Mhine Intelligene, 005. 7(8): p. 165-178. 33. Csurk, G., et l., Visul tegoriztion with bg of keypoints, in ECCV '04. 004. p. 1-. 34. Lowe, D.G., Towrds omputtionl model for objet reognition in IT ortex, in Biologilly Motivted Computer Vision, Proeeding. 000. p. 0-31. 35. Lowe, D.G., Distintive imge fetures from sle-invrint keypoints. Interntionl ournl of Computer Vision, 004. 60(): p. 91-110.