Neural Network-Based Detection for Faces in Cluttered Scenes

Abstract

Computer- based technol og i es a n d tool s ror industri a l mach i l1 e ,•i sion (111d sci cnti lie im age processing a re qui ckl y being incorporated in to a \'clr iety or appl i cntions. The goa l o r m achin e v i s i on i s to process im ages acq u ired by ca meras 111 ord er to produce a n appropri a te represe nt ati on o r the objec ts i n t he rea l world. The goal o r thi s th es i s i s t o crea t e a computer i zed me t hod ror automati c 1l cc d etecti on in c luttered scenes. Th e m eth od we u sed i s ge neral a nd ca n be ap pli ed to m a n y object d etecti on probl ems with l i ttl e mod i lieations. ln thi s thes is, we present a neura l nelvvor k- bascd u pri ght rront al 1l ce detecti on system . Gen era ll y, object d et ecti on i s th e r robl e m or det er m in i ng w hether or n ot a sub-\,v ind ow o r a n ima ge bel ongs to the se t o f ' ima ges o r (Ill o bject or i11terest. Th e sys te m task i s to d et ect a nd l oca te upright 1•ru,1 tal human races. 111 a graysca l e image tha t co nt a ins h uma n l1 ces against cl uttered ba ckground . The sys tem.•proposes so me soluti ons to the prob l ems rel a ted to the IJCe d etecti on d oma in . It a r bitrates bet ween multipl e n e ural ne t wor k s and

heuristi cs. su ch as the

1l et th at races ra rel y overlap in im ages. t o im pro\'l' the

pcrrorma nce and accuracy o r the used a l gorithm. We used a nc,,• pre limin a ry a l gorithm t o increase the speed or th e syst em t o 3 - I 0 t i m es 1l st e r than oth er sys tems. Thi s new a l gorithm ex tracts a ll the 20X20pi xe l s su b-w ind ows th a t ha ve hi gh possibility or conta in i n g races. a nd process the m by the system . The system as " w h ol e i s co mposed o f' th e f oll ow ing t wo s t ages: (i } The

pre processiug stage. in w hi ch the system app li es " set o r fi lt e rs to the i111a gc to . increase the d et ecti on s peed a nd reduce the variati on caused b' l i gh ti11g or camera di llerenccs. (il) The Neural Network Det ection Stu f!_e , '' hich i s composed or three d etecti on neural net works. The sys tem a rbitra t es between the output o r these three ne twor k s t o get the best pe r rorm a nce.

We used cli[Terent a rbitra ti on and heuristics t o implem ent diffe rent syst em s. These system s were test ed on 77 dirrerent test images containin g a t ot a l o f' 277 l1ces. Th ese systems were ab l e t o detect bet ween 83. I % a nd 94.(1n1> o f' laces i n a ll im ages with a n accepta bl e number or fal se d etection s. Th e prope r sys tem to be u sed d e pends on th e a pplica ti on th 3l thi s syst e m will be used i n.