Wake wahlabeka umxhwele ngekhono lekhamera ye-smartphone yakho ukubona ubuso esithombeni seqembu?
Mhlawumbe uye wamangazwa indlela izimoto ezizishayelayo ezihamba ngayo ngaphandle komthungo, zikhomba abahamba ngezinyawo nezinye izimoto ngokunemba okuyisimangaliso.
Lokhu okufeziwe okubonakala kungaphezu kwemvelo kwenziwa ukuba kwenzeke ngokutholwa kwento, isihloko esithakazelisayo socwaningo. Kalula nje, ukutholwa kwento ukuhlonza kanye nokwenza kwasendaweni kwezinto ngaphakathi kwezithombe noma amavidiyo.
Ubuchwepheshe obuvumela amakhompyutha ukuthi “abone” futhi aqonde umhlaba owazungezile.
Kodwa isebenza kanjani le nqubo emangalisayo? Siyakubona lokho ukufunda okujulile kune iguqule indawo yokuhlonza into. Ivula indlela yezinhlelo zokusebenza eziningi ezinethonya eliqondile ezimpilweni zethu zansuku zonke.
Kulokhu okuthunyelwe, sizodlula endaweni ethokozisayo yokuhlonza into esekelwe ekufundeni okujulile, sifunde ukuthi inamandla okubumba kabusha indlela esisebenzisana ngayo nobuchwepheshe.
Kuyini Kahle Ukutholwa Kwezinto?
Okukodwa umbono wekhompyutha oyisisekelo imisebenzi ukutholwa kwento, okubandakanya ukuthola nokuthola izinto ezahlukahlukene esithombeni noma kuvidiyo.
Uma kuqhathaniswa nokuhlukaniswa kwesithombe, lapho kunqunywa ilebula yesigaba sento ngayinye, ukutholwa kwento kuya ngesinyathelo esisodwa ngokuqhubekayo ngokungagcini nje ngokukhomba ubukhona bento ngayinye kodwa futhi nokudweba amabhokisi abophayo azungeze ngalinye.
Njengomphumela, singakwazi ukuhlonza ngesikhathi esisodwa izinhlobo zezinto esizithandayo futhi sizibeke ngokunembile.
Ikhono lokubona izinto libalulekile ezinhlelweni eziningi, kuhlanganisa ukushayela okuzimele, ukugadwa, ukubonwa kobuso, kanye nezithombe zezokwelapha.
Ukuze kusingathwe le nselele enzima ngokunemba okuvelele nokusebenza kwesikhathi sangempela, amasu asekelwe ekufundeni aguqule ukutholwa kwezinto.
Ukufunda okujulile kusanda kuvela njengesu elinamandla lokunqoba lobu bunzima, ukushintsha imboni yokuqaphela izinto.
Umndeni wakwaR-CNN kanye ne YOLO umndeni imindeni emibili eyimodeli eyaziwayo ekuhlonzeni into ezocutshungulwa kulesi sihloko.
Umndeni wakwa-R-CNN: Ukutholwa Kwento Yokuphayona
Ucwaningo lokuqala lokuqashelwa kwento lubone intuthuko enkulu ngenxa yomndeni we-R-CNN, ohlanganisa i-R-CNN, i-Fast R-CNN, ne-Faster R-CNN.
Ngokwakheka kwayo okunamamojula amathathu, izifunda ezihlongozwayo ze-R-CNN zisebenzisa i-CNN ukuze zikhiphe izici, nezinto ezihlukanisiwe zisebenzisa ama-SVM aqondile.
I-R-CNN yayilungile, nakuba kuthathe isikhathi ngoba amabhidi esifunda sekhandidethi ayedingeka. Lokhu kwasingathwa yi-Fast R-CNN, ekhulise ukusebenza kahle ngokuhlanganisa wonke amamojula abe yimodeli eyodwa.
Ngokungeza i-Regional Proposal Network (RPN) edale futhi yathuthukisa iziphakamiso zesifunda phakathi nokuqeqeshwa, i-R-CNN esheshayo ithuthukise kakhulu ukusebenza futhi yazuza cishe ukubonwa kwento yesikhathi sangempela.
Ukusuka ku-R-CNN kuya ku-Faster R-CNN
Umndeni wakwa-R-CNN, okusho ukuthi “Region-based I-Convolutional Neural Networks,” ithuthuke ekutholweni kwento.
Lo mndeni uhlanganisa i-R-CNN, i-Fast R-CNN, ne-Faster R-CNN, zonke eziklanyelwe ukubhekana nemisebenzi yokwenziwa kwasendaweni nemisebenzi yokuqaphela.
I-R-CNN yasekuqaleni, eyethulwe ngo-2014, yabonisa ukusetshenziswa ngempumelelo kwamanethiwekhi e-convolutional neural ukuze kutholwe into nokwenziwa kwendawo.
Kuthathe isu lezinyathelo ezintathu elihlanganisa isiphakamiso sesifunda, ukukhishwa kwesici nge-CNN, kanye nokuhlukaniswa kwento ngezihlukanisi ezilandelanayo zomshini Wokusekela Vector (SVM).
Ngemva kokwethulwa kwe-Fast R-CNN ngo-2015, izinkinga zejubane zaxazululwa ngokuhlanganisa isiphakamiso sesifunda nokuhlukaniswa kube imodeli eyodwa, okwehliswa kakhulu ukuqeqeshwa nesikhathi sokunquma.
I-Faster R-CNN, ekhishwe ku-2016, ithuthukise isivinini nokunemba ngokufaka i-Regional Proposal Network (RPN) ngesikhathi sokuqeqeshwa ukuphakamisa ngokushesha nokubuyekeza izindawo.
Njengomphumela, i-Faster R-CNN isizisungule njengenye yama-algorithms ahamba phambili emisebenzi yokuthola into.
Ukufakwa kwezihlukanisi zezigaba ze-SVM bekubalulekile empumelelweni yomndeni we-R-CNN, ukushintsha indawo yokubona ikhompuyutha nokubeka indlela yezimpumelelo zesikhathi esizayo ekutholweni kwezinto ezisekelwe ekufundeni okujulile.
Amandla:
- Ukunemba kokutholwa kwento yokwenza kwasendaweni okuphezulu.
- Ukunemba nokusebenza kahle kulinganisa idizayini ehlanganisiwe ye-R-CNN esheshayo.
Ukubuthakathaka:
- Ukuthintwa nge-R-CNN kanye ne-Fast R-CNN kungaba nzima kakhulu.
- Ukuze i-R-CNN esheshayo isebenze ngokusemandleni ayo, iziphakamiso eziningi zesifunda zingase zisadingeka.
Umndeni we-YOLO: Ukutholwa Kwento Ngesikhathi Sangempela
Umndeni we-YOLO, osuselwe kumqondo othi “Ubukeka Kanye Kuphela” ugcizelela ukubonwa kwento yesikhathi sangempela kuyilapho udela ukunemba.
Imodeli yoqobo ye-YOLO yayihlanganisa inethiwekhi eyodwa ye-neural eyabikezela ngokuqondile amabhokisi abophayo namalebula ekilasi.
Naphezu kokuba nokunemba okuncane kokubikezela, i-YOLO ingasebenza ngesivinini esingafika kumafreyimu angu-155 ngesekhondi. I-YOLOv2, eyaziwa nangokuthi yi-YOLO9000, yabhekana nokunye ukushiyeka kwemodeli yasekuqaleni ngokubikezela izigaba zezinto ezingu-9,000 futhi ifaka namabhokisi okusetshenzelwa phansi ukuze uthole izibikezelo eziqinile.
I-YOLOv3 ithuthuke nakakhulu, ngenethiwekhi yesitholi sesici esibanzi.
Ukusebenza Kwangaphakathi Komndeni we-YOLO
Amamodeli okuhlonza into emndenini we-YOLO (Ubukeka Kanye Kuphela) avele njengempumelelo ephawulekayo ekuboneni ngekhompyutha.
I-YOLO, eyethulwa ngo-2015, ibeka phambili isivinini nokuhlonza into yesikhathi sangempela ngokulindela ngokuqondile amabhokisi abophezelayo namalebula ekilasi.
Nakuba ukunemba okuthile kuhlatshelwe, ihlaziya izithombe ngesikhathi sangempela, ikwenze kube usizo ezinhlelweni ezibucayi zesikhathi.
I-YOLOv2 ihlanganise amabhokisi okubamba ihange okubhekana nezikali zezinto ezihlukahlukene futhi yaqeqeshwa kumadathasethi amaningi ukuze ilindele amakilasi ezinto angaphezu kuka-9,000.
Ngo-2018, i-YOLOv3 yathuthukisa umndeni nakakhulu ngenethiwekhi yesitholi sesici esijulile, ethuthukisa ukunemba ngaphandle kokudela ukusebenza.
Umndeni we-YOLO ubikezela amabhokisi ahlanganisayo, amathuba ekilasi, kanye nezikolo zokungaboni ngaso linye ngokuhlukanisa isithombe sibe yigridi. Ihlanganisa kahle isivinini nokunemba, iyenze ivumelane nezimo ukuze isetshenziswe kuyo izimoto ezizimele, ukugadwa, ukunakekelwa kwezempilo, neminye imikhakha.
Uchungechunge lwe-YOLO luguqule ukuhlonza into ngokunikeza izixazululo zesikhathi sangempela ngaphandle kokudela ukunemba okubalulekile.
Kusukela ku-YOLO kuya ku-YOLOv2 kanye ne-YOLOv3, lo mndeni wenze intuthuko enkulu ekuthuthukiseni ukuqashelwa kwento kuzo zonke izimboni, usungula indinganiso yezinhlelo zesimanje zokutholwa kwezinto ezisekelwe ekufundeni okujulile.
Amandla:
- Ithola izinto ngesikhathi sangempela ngamazinga aphezulu ozimele.
- Ukuzinza ekubikezelweni kwebhokisi elibophezelayo kwethulwa ku-YOLOv2 naku-YOLOv3.
Ukubuthakathaka:
- Amamodeli e-YOLO angayeka ukunemba okuthile ngokushintshanisa isivinini.
Ukuqhathanisa Komndeni Okuyisibonelo: Ukunemba vs. Ukusebenza kahle
Uma imindeni ye-R-CNN kanye ne-YOLO iqhathaniswa, kuyacaca ukuthi ukunemba nokusebenza kahle kuwukuhwebelana okubalulekile. Amamodeli omndeni we-R-CNN ahamba phambili ngokunemba kodwa ayanensa ngesikhathi sokucatshangelwa ngenxa yokwakheka kwawo okunamamojula amathathu.
Umndeni we-YOLO, ngakolunye uhlangothi, ubeka phambili ukusebenza kwesikhathi sangempela, uhlinzeka ngesivinini esivelele ngenkathi ulahlekelwa ukunemba okuthile. Isinqumo phakathi kwale mindeni eyimodeli sinqunywa izidingo eziqondile zohlelo lokusebenza.
Amamodeli omndeni we-R-CNN angase akhetheke emsebenzini omningi odinga ukunemba ngokwedlulele, kuyilapho amamodeli omndeni we-YOLO elungele izinhlelo zokusebenza zesikhathi sangempela.
Ngaphandle Kokubonwa Kwento: Izicelo Zomhlaba Wangempela
Ngale kwemisebenzi ejwayelekile yokuqaphela into, ukutholwa kwento okusekelwe ekufundeni okujulile kuthole ukusetshenziswa okubanzi.
Ukuvumelana nezimo nokunemba kwayo kudale amathuba amasha emikhakheni eyahlukene, ukubhekana nezinselele eziyinkimbinkimbi nokuguqula amabhizinisi.
Izimoto Ezizimelayo: Ukusetha Izinga Lokushayela Okuphephile
Ukutholwa kwento kubalulekile ezimotweni ezizimele ukuze kuqinisekiswe ukuzulazula okuphephile nokuthembekile.
Amamodeli wokufunda okujulile hlinzeka ngolwazi olubalulekile lwezinhlelo zokushayela ezizimele ngokubona kanye nokwenza indawo yabahamba ngezinyawo, abagibeli bamabhayisikili, ezinye izimoto, kanye nezingozi zomgwaqo ezingaba khona.
Lezi zinhlobo zivumela izimoto ukuthi zithathe izinqumo zesikhathi sangempela futhi zigweme ukushayisana, okusisondeza ekusaseni lapho izimoto ezizishayelayo zihlala khona kanye nabashayeli abangabantu.
Ukwandisa Ukusebenza Nokuvikeleka Embonini Yokudayisa
Ibhizinisi lokudayisa lamukele ukutholwa kwento ejulile esekelwe ekufundeni ukuze kuthuthukiswe kakhulu ukusebenza kwalo.
Izinsiza zokuthola into ekuhlonzweni nasekulandeleni imikhiqizo emashalofini esitolo, okuvumela ukuphinda kusetshenziswe ngempumelelo kanye nokwehliswa kwezimo zokungabi khona kwesitoko.
Ngaphezu kwalokho, amasistimu okugada afakwe ama-algorithms okuthola izinto asiza ekuvimbeleni ukweba kanye nokugcinwa kokuphepha kwesitolo.
Ukuthuthukiswa Kwezithombe Zezokwelapha Ekunakekelweni Kwezempilo
Ukutholwa kwezinto ezisekelwe ekufundeni okujulile sekuyithuluzi elibalulekile ekucabangeni kwezokwelapha emkhakheni wezokunakekelwa kwempilo.
Isiza abasebenzi bezempilo ukuthi babone okungavamile kuma-X ray, izikena ze-MRI, nezinye izithombe zezokwelapha, ezinjengomdlavuza noma ukukhubazeka.
Izinsiza zokuhlonza izinto ekuxilongweni kusenesikhathi nasekuhleleni ukwelashwa ngokuhlonza nokugqamisa izindawo ezithile ezikhathazayo.
Ukuthuthukisa Ukuphepha Ngokuvikeleka Nokugada
Ukutholwa kwento kungaba usizo ngendlela emangalisayo ekuvikelekeni nasekugadweni izinhlelo zokusebenza.
Ama-algorithms wokufunda okujulile ukusiza izixuku zamawashi, ukuhlonza ukuziphatha okusolisayo, nokuthola izingozi ezingaba khona ezindaweni zomphakathi, ezikhumulweni zezindiza, nasezindaweni zokuhamba.
Lezi zinhlelo zingaxwayisa ochwepheshe bezokuphepha ngesikhathi sangempela ngokuqhubeka nokuhlola okuphakelayo kwamavidiyo, ukuvimbela ukwephulwa kwezokuphepha, nokuqinisekisa ukuphepha komphakathi.
Izithiyo Zamanje kanye Namathemba Esikhathi esizayo
Naphezu kwenqubekelaphambili ebalulekile ekutholweni kwezinto ezisekelwe ekufundeni okujulile, izinkinga zisekhona. Ubumfihlo bedatha buyinkinga enkulu, njengoba ukutholwa kwento ngokuvamile kuhlanganisa nokuphatha ulwazi olubucayi.
Enye inkinga ebalulekile ukuqinisekisa ukuqina ekuhlaselweni yizitha.
Abacwaningi basafuna izindlela zokwandisa ukwenziwa kwemodeli nokuchazwa.
Ngocwaningo oluqhubekayo olugxile ekuhlonzeni izinto eziningi, ukulandela umkhondo wezinto zevidiyo, kanye nokubonwa kwento ye-3D yesikhathi sangempela, ikusasa libonakala liqhakazile.
Kufanele silindele izixazululo ezinembe nakakhulu nezisebenza kahle maduzane njengoba amamodeli okufunda okujulile eqhubeka nokukhula.
Isiphetho
Ukufunda okujulile kuguqule ukutholwa kwezinto, kwangenisa inkathi yokunemba okukhulu nokusebenza kahle. Imindeni ye-R-CNN kanye ne-YOLO idlale indima ebalulekile, ngayinye inekhono elihlukile lezinhlelo zokusebenza ezithile.
Ukuhlonza into esekelwe ekufundeni okujulile kuguqula imikhakha futhi kuthuthukiswe ukuphepha nokusebenza kahle, ukusuka ezimotweni ezizimele kuye ekunakekelweni kwezempilo.
Ikusasa lokutholwa kwento libonakala liqhakazile kunangaphambili njengoba ucwaningo luthuthuka, kubhekwana nobunzima kanye nokuhlola izindawo ezintsha.
Sibona ukuzalwa kwenkathi entsha ekuboneni ngekhompyutha njengoba samukela amandla okufunda okujulile, nokutholwa kwezinto okuhamba phambili.
shiya impendulo