RIT Real-time and Image Processing Lab 83 Lomb Memorial Drive, Rocester, !" #$%&3 aa'()&*rit+edu 1. Overview ,lassi-ication o- images based on scenes .resents te di--icult tas/ o- identi-0ing -eatures s.eci-ic to te scene+ Te 1ide range o- scenes, suc as urban, countr0, mountains, -orests, etc+, ma0 introduce -eatures, 1ic ma0 com.licate te classi-ication .rocess+ 2iven te limited time -rame to com.lete te .ro3ect, te range 1ill be condensed to urban, suburban, and countr0 scenes+ Te goal o- tis .ro3ect 1ill be to correctl0 classi-0 eac image in a test set based on t0.e o- scene+ &+ Background A sim.le a..roac ins.ired b0 Pa0ne and 4ing 5$6 uses edge -eatures to determine te straigtness metric, 1ic measures te .ercentage o- straigt edges+ Tis algoritm 1as intended -or classi-0ing indoor and outdoor images, o1ever, tis ma0 .ave te 1a0 to se.arate te tree t0.es o- scenes+ 7or e8am.le, urban scenes -illed buildings 1ill .roduce more straigt edges tan a countr0 scene+ Anoter a..roac to image classi-ication -irst .roduces models based on a training set o- /no1n classi-ication+ 9osc, :isserman, Muno; 5#6 introduced .robabilistic Latent 4emantic Anal0sis <.L4A= modeling to classi-0 di--erent scenes+ A-ter models are generated, te test images are com.ared to te models and classi-ied based on te similarit0 o- -eatures+ Te scene classi-ication using .L4A modeling 1as derived -rom te more commonl0 /no1n bag o- -eatures <or 1ords= metod develo.ed b0 7ei 7ei+ Additional researc conducted b0 Mars;ale/ 5&6 uses a sim.le .rocedure -or image classi-ication+ 7irst, interest .oints are detected using a La.lacian detector, 1ic detects blobs+ Local descri.tors are -ormed using te interest .oints using te 4I7T metod+ 7rom te descri.tors, te bag o- -eatures metod com.ares -eatures o- te test images to tose o- te training set+ Te test image is classi-ied into te best matc training set+ Also using a test set to .roduce a model, Li.son, 2rimson, and 4ina 536 divides te images into signi-icant regions+ Te relationsi. bet1een te regions, suc as distance, orientation and color, are encoded into te basic model o- te scene+ Test images are divided into regions o- interest and com.ared to e8isting models to be classi-ied+ 4ource code 1ill be -rom scratc using e8tracted algoritms -rom te re-erenced .a.ers and in-class discovered tecni>ues+ Tese original .a.ers 1ere -ound using te 2oogle 4colar searc engine+ Te source code 1ill be 1ritten as MATLA9 scri.ts and -unctions and e8ecuted using MATLA9 so-t1are+ 3+ Progress 4everal algoritms and metods to classi-0 images into .redetermined categories ave been researced+ Most algoritms re>uire a .reviousl0 identi-ied training set o- images, so sam.le images 1ere collected -rom te Internet using a sim.le image searc engine+ To ma/e te current .ro3ect direction clear, sim.le testing o- eac metod as been conducted to determine te >uic/est and most accurate 1a0 o- classi-0ing images+ Te most accurate metod a..ears to be te bag o- -eatures metod+ Ho1ever, tis metod ma0 .rove to be te most com.le8, re>uiring te greatest number o- .rocedures+ It migt be bene-icial to combine one or t1o metods to sim.li-0 te .rocedure, suc as -iltering out images 1it a large .ercentage o- straigt edges, 1ic 1ould surel0 be classi-ied as an urban scene+ Te -irst -e1 ste.s o- te 0brid metod as been im.lemented using a -e1 di--erent tecni>ues+ ?dge detection algoritms 1ere im.lemented using te built-in MATLA9 edge -unction and several di--erent edge detectors, suc as La.lacian o- 2aussian, 4obel, and ,ann0+ 4traigt edges 1ere detected using te ouglines -unction+ Tis resulted in de-ined straigt edges o- te training images+ In addition, a 4I7T toolbo8 -or MATLA9 1as located and installed, because 4I7T im.lementation is not 1itin te sco.e o- te .ro3ect+ Te 4I7T tool 1as used to de-ine -eatures and descri.tors o- te training images+ Tese descri.tors 1ill be used to develo. te visual 1ords -or eac categor0+ $+ Data Test images 1ill be .rovided b0 te RIT Introduction to ,om.uter @ision course 1eb .age+ Ho1ever, i- a training set and model is necessar0 -or te im.lementation o- te .ro3ect, tis data 1ill be collected e8ternall0 troug te 1eb+ Anl0 images in te -ormat BP?2 1ill be collected and tested+ In addition, results -rom .revious researc 1ill be e8tracted -rom te re-erenced .a.er and com.ared to te im.lemented .ro3ect+ 5. eferences 5#6 9osc, Anna, Andre1 :isserman, and Cavier Muno;+ D4cene ,lassi-ication via PL4A+D Proc. ECCV <&EE%=F (#)-3E+ CiteULike+ Geb+ #3 Ban+ &E##+ Htt.FII111+citeuli/e+orgIuserIina0at/IarticleI3')'#)EJ+ 5&6 Mars;aKe/, Marcin, Bianguo :ang, 4vetlana La;ebni/, and ,ordelia 4cmid+ D9ag-o-- -eatures Image Re.resentationF 4tate-o--te-art and 9e0on+D BournLe Du 2DR I4I4, #( 4e.t+ &EE%+ Geb+ #% Ban+ &E##+ Htt.FII111-.oleia+li.%+-rIMcordIisisImars;ale/+.d-J+ 536 P+ Li.son, ?+ 2rimson, P+ 4ina, D,on-iguration based scene classi-ication and image inde8ing,D cv.r, ..+#EE), #'') I??? ,om.uter 4ociet0 ,on-erence on ,om.uter @ision and Pattern Recognition <,@PRN')=, #'')+ 5$6 Pa0ne, Andre1, and 4ameer 4ing+ DIndoor vs+ Autdoor 4cene ,lassi-ication in Digital Potogra.s+D Pattern Recognition 38+#E <&EE(=F #(33-($(+ ScienceDirect - Home+ ?lsevier 9+@+ Geb+ #3 Ban+ &E##+ Htt.FII111+sciencedirect+comIscienceO PobQArticleRRLSPudiQ9%@#$-$7C&3D$- $SPuserQ$)EE$SPcoverDateQ#EI3#I&EE(SPrdocQ#SP-mtQigSPorigQsearcSPoriginQs earcSPsortQdSPdocancorQSvie1QcSPsearc4trIdQ#%E%&3&3'ESPrerunAriginQgoogle SPacctQ,EEEEE(E#8SPversionQ#SPurl@ersionQESPuseridQ$)EE$Smd(Qee$'E3d%'%3')a #'cd38b)$$)$e&)(e(Ssearct0.eQaJ+