Kuyini i-Computer Vision ku-AI?

Uma wake wavula ifoni yakho ngobuso bakho, waskena irisidi, noma wabuka ikhamera yokuzihlola uzibuza ukuthi ingabe iyahlulela i-avocado yakho, uphambene nombono wekhompyutha. Kalula nje, i-Computer Vision ku-AI yindlela imishini efunda ngayo ukubona nokuqonda kuyamangaza ? Futhi yebo. Futhi ngezinye izikhathi kuyathusa uma sithembekile. Uma kukuhle kakhulu, iguqula amaphikseli angahlelekile abe yizenzo ezisebenzayo. Uma kukubi kakhulu, iyaqagela futhi iyanyakaza. Ake singene kahle.

Izihloko ongase uthande ukuzifunda ngemva kwalesi:

🔗 Kuyini ukucwasa kwe-AI
Indlela ukucwasa okwakheka ngayo ezinhlelweni ze-AI nezindlela zokukuthola nokuyinciphisa.

🔗 Kuyini i-AI yokubikezela
Indlela i-AI yokubikezela esebenzisa ngayo idatha ukubikezela izitayela nemiphumela.

🔗 Uyini umqeqeshi we-AI?
Imithwalo yemfanelo, amakhono, namathuluzi asetshenziswa ochwepheshe abaqeqesha i-AI.

🔗 Iyini i-Google Vertex AI
Ukubuka Konke kwepulatifomu ye-AI ehlanganisiwe ye-Google yokwakha nokusabalalisa amamodeli.

Kuyini i-Computer Vision ku-AI, ngempela? 📸

I-Computer Vision ku-AI iyigatsha lobuhlakani bokwenziwa elifundisa amakhompyutha ukuhumusha nokucabanga ngedatha ebonakalayo. Yindlela esukela kuma-pixel angavuthiwe kuya encazelweni ehlelekile: “lesi yisibonakaliso sokuma,” “labo bangabahamba ngezinyawo,” “ukushintshwa kwe-weld kunephutha,” “inani le-invoyisi selifikile.” Ihlanganisa imisebenzi efana nokuhlukaniswa, ukuthola, ukuhlukaniswa, ukulandelela, ukulinganisa ukujula, i-OCR, nokunye okuhlanganiswe ndawonye ngamamodeli okufunda amaphethini. Insimu esemthethweni ihlanganisa i-geometry yakudala kuya ekufundeni okujulile kwesimanje, ngezincwadi zokudlala ezisebenzayo ongazikopisha futhi uzilungise. [1]

Indaba esheshayo: cabanga ngomugqa wokupakisha onekhamera encane engu-720p. I-detector elula ibona ama-caps, bese i-tracker elula iqinisekisa ukuthi zibekwe ozimele abahlanu abalandelanayo ngaphambi kokukhanyisa ibhodlela ngokukhanya okuluhlaza. Akuyona into enhle kodwa eshibhile, iyashesha, futhi inciphisa ukuhlelwa kabusha.

Yini eyenza i-Computer Vision ku-AI ibe wusizo? ✅

• Ukugeleza kwesignali kuya esenzweni : Okokufaka okubonakalayo kuba umphumela ongasetshenziswa. Ideshibhodi encane, izinqumo eziningi.

• Ukuhlanganisa : Ngedatha efanele, imodeli eyodwa iphatha izinhlobo ezahlukene zezithombe. Akuphelele - ngezinye izikhathi kuyamangaza.

• Ukusetshenziswa kwedatha : Amakhamera ashibhile futhi yonke indawo. Umbono uguqula lolo lwandle lwamaphikseli lube ukuqonda.

• Isivinini : Amamodeli angacubungula ozimele ngesikhathi sangempela kuhadiwe encane - noma cishe ngesikhathi sangempela, kuye ngomsebenzi kanye nesisombululo.

• Ukuhlanganiswa : Hlanganisa izinyathelo ezilula zibe izinhlelo ezithembekile: ukutholwa → ukulandelela → ukulawulwa kwekhwalithi.

• Uhlelo lwe-Eco : Amathuluzi, amamodeli aqeqeshwe kusengaphambili, izilinganiso, kanye nokusekelwa komphakathi - imakethe enkulu yekhodi.

Masibe neqiniso, imfihlo ayiyona imfihlo: idatha enhle, ukuhlolwa okuhlelekile, ukuqaliswa ngokucophelela. Okunye ukuzijwayeza... futhi mhlawumbe ikhofi. ☕

Indlela i-Computer Vision esebenza ngayo ku-AI, ngendlela enengqondo 🧪

1. Ukutholwa kwesithombe
   Amakhamera, amaskena, ama-drone, amafoni. Khetha uhlobo lwenzwa, ukuvezwa, ilensi, kanye nesivinini sohlaka ngokucophelela. Faka udoti ngaphakathi, njll.

2. Ukucubungula Kwangaphambili
   Shintsha usayizi, nquma, lungisa, susa umsindo noma ususe umsindo uma kudingeka. Ngezinye izikhathi ukushintsha okuncane komehluko kususa izintaba. [4]

3. Amalebula namasethi edatha
   Amabhokisi ahlanganisiwe, ama-polygons, amaphuzu ayisihluthulelo, ama-span ombhalo. Amalebula alinganiselayo, amele - noma imodeli yakho ifunda imikhuba ephambene.

4. Ukumodela

   • Ukuhlukaniswa : “Yisiphi isigaba?”

   • Ukutholwa : "Ziphi izinto?"

   • Ukuhlukaniswa : “Yimaphi amaphikseli angawayiphi into?”

   • Amaphuzu ayisihluthulelo kanye nokuma : “Ziphi izindawo zokuhlangana noma izindawo ezibalulekile?”

   • I-OCR : “Yimuphi umbhalo osesithombeni?”

   • Ukujula kanye ne-3D : “Konke kukude kangakanani?”
     Izakhiwo ziyahlukahluka, kodwa amanethi e-convolutional kanye namamodeli esitayela se-transformer ayabusa. [1]

5. Ukuqeqesha
   Hlukanisa idatha, lungisa ama-hyperparameter, uhlele kabusha, ungeze. Ukuma kusenesikhathi ngaphambi kokuba ubambe ngekhanda iphepha lodonga.

6. Ukuhlola
   Sebenzisa izilinganiso ezifanele umsebenzi njenge-mAP, IoU, F1, CER/WER ye-OCR. Ungakhethi kahle. Qhathanisa kahle. [3]

7. Ukufakwa
   Lungiselela okuqondiwe: imisebenzi ye-cloud batch, ukuqagela kudivayisi, amaseva onqenqemeni. Qapha ukuzulazula. Ziqeqeshe kabusha lapho umhlaba ushintsha.

Amanethi ajulile akhuthaze ukukhuphuka kwekhwalithi lapho amasethi edatha amakhulu kanye nokubala kufinyelela ku-critical mass. Izimpawu ezifana nenselele ye-ImageNet zenze leyo nqubekela phambili yabonakala - futhi yangapheli. [2]

Imisebenzi eyinhloko ozoyisebenzisa ngempela (futhi nini) 🧩

• Ukuhlukaniswa kwesithombe : Ilebula elilodwa ngesithombe ngasinye. Sebenzisa izihlungi ezisheshayo, ukuhlunga, noma amasango ekhwalithi.

• Ukutholwa kwezinto : Amabhokisi azungeze izinto. Ukuvimbela ukulahleka kwezitolo, ukutholwa kwezimoto, inani lezilwane zasendle.

• Ukuhlukaniswa kwesimo : Imidwebo enembile njengephikseli ngento ngayinye. Amaphutha okukhiqiza, amathuluzi okuhlinza, i-agritech.

• Ukuhlukaniswa kwe-Semantic : Isigaba ngephikseli ngaphandle kokuhlukanisa izimo. Izigcawu zemigwaqo yasemadolobheni, isembozo sezwe.

• Ukutholwa kwe-Keypoint kanye nokuma : Amalunga, izimpawu ezibalulekile, izici zobuso. Ukuhlaziywa kwezemidlalo, i-ergonomics, i-AR.

• Ukulandelela : Landela izinto ngokuhamba kwesikhathi. Ukuthutha, ithrafikhi, ukuphepha.

• I-OCR kanye ne-AI yedokhumenti : Ukukhishwa kombhalo kanye nokuhlaziywa kwesakhiwo. Ama-invoyisi, amarisidi, amafomu.

• Ukujula kanye ne-3D : Ukwakhiwa kabusha kusuka ekubukeni okuningi noma ezibonisweni ezi-monocular. Amarobhothi, i-AR, imephu.

• Ukubhala amagama angezansi : Fingqa izigcawu ngolimi lwemvelo. Ukufinyeleleka, usesho.

• Amamodeli olimi lombono : Ukucabanga okunezindlela eziningi, umbono okhuliswe ngokubuyisa, i-QA eyisisekelo.

Isimo esincane sekesi: ezitolo, i-detector ikhomba ukuthi ayikho indawo ebheke eshalofini; i-tracker ivimbela ukubalwa kabili njengoba abasebenzi bephinda bafake izinto; umthetho olula uqondisa ozimele abangazethembi ekubuyekezweni ngabantu. Yiqembu elincane le-orchestra elihlala lihambisana.

Ithebula lokuqhathanisa: amathuluzi okuthumela ngokushesha 🧰

Kuyinqaba ngamabomu. Yebo, izikhala ziyamangaza - ngiyazi.

Yini eqhuba ikhwalithi ku- Computer Vision ku-AI 🔧

• Ukumbozwa kwedatha : Ukushintsha kokukhanya, ama-engeli, izizinda, ama-edge cases. Uma kungenzeka, kufake.

• Ikhwalithi yelebula : Amabhokisi angaguquguquki noma ama-polygons angahlelekile ayasabisa i-mAP. I-QA encane ihamba ibanga elide.

• Ukwandiswa okuhlakaniphile : Nciphisa, zungezisa, shintsha ukukhanya, engeza umsindo wokwenziwa. Yiba ngokoqobo, kungabi yisiphithiphithi esingahleliwe.

• Ukufaneleka kokukhetha imodeli : Sebenzisa ukutholwa lapho kudingeka ukutholwa khona-ungaphoqi umuntu ohlukanisayo ukuthi aqagele izindawo.

• Izilinganiso ezihambisana nomthelela : Uma imiphumela emibi engamanga ilimaza kakhulu, lungisa ukukhumbula. Uma imiphumela emibi engamanga ilimaza kakhulu, qale ngokunemba.

• I-loop yempendulo eqinile : Ukwehluleka kwelogi, ukulebula kabusha, ukuqeqesha kabusha. Hlanza, phinda. Kuyisicefe kancane-kusebenza kahle kakhulu.

Ukuze kutholakale/kuhlukaniswe, indinganiso yomphakathi i- Average Precision elinganiselwe kuwo wonke ama-IoU thresholds-aka COCO-style mAP . Ukwazi ukuthi i-IoU kanye ne-AP@{0.5:0.95} kubalwa kanjani kugcina izimangalo zebhodi yabaphambili zingakumangazi ngamadesimali. [3]

Amacala okusetshenziswa kwangempela angewona acatshangelwayo 🌍

• Ukuthengisa : Ukuhlaziya ishelufu, ukuvimbela ukulahleka, ukuqapha umugqa, ukuhambisana ne-planogram.

• Ukukhiqiza : Ukutholwa kwesici somphezulu, ukuqinisekiswa kwenhlangano, isiqondiso serobhothi.

• Ukunakekelwa Kwezempilo : Ukuhlolwa kwe-X-ray, ukutholwa kwezinsimbi, ukuhlukaniswa kwamaseli.

• Ukuhamba : I-ADAS, amakhamera omgwaqo, ukupaka abantu, ukulandelela ukuhamba kwe-micromobility.

• Ezolimo : Ukubalwa kwezitshalo, ukubona izifo, ukulungela isivuno.

• Umshuwalense Nezezimali : Ukuhlolwa komonakalo, ukuhlolwa kwe-KYC, amafulegi okukhwabanisa.

• Ukwakha kanye Namandla : Ukuthobela imithetho yokuphepha, ukutholwa kokuvuza, ukuqapha ukugqwala.

• Okuqukethwe Nokufinyeleleka : Amagama-ncazo azenzakalelayo, ukulinganisela, ukusesha okubonakalayo.

Iphethini ozoyiqaphela: shintsha ukuskena ngesandla nge-othomathikhi, bese ikhuphukela kubantu lapho ukuzethemba kwehla. Akuyona into ekhangayo - kodwa iyakhula.

Idatha, amalebula, kanye nezilinganiso ezibalulekile 📊

• Ukuhlukaniswa : Ukunemba, i-F1 yokungalingani.

• Ukutholwa : i-mAP ngaphesheya kwemingcele ye-IoU; hlola amabhakede e-AP kanye nosayizi ngamunye. [3]

• Ukuhlukaniswa : mIoU, Idayisi; hlola namaphutha ezingeni lesibonelo.

• Ukulandelela : MOTA, IDF1; ikhwalithi yokuhlonza kabusha iyiqhawe elithule.

• I-OCR : Izinga Lephutha Lohlamvu (CER) kanye Nezinga Lephutha Legama (WER); ukwehluleka kwesakhiwo kuvame ukuba yinto eyinhloko.

• Imisebenzi yokubuyela emuva : Ukujula noma ukuma kusebenzisa amaphutha aphelele/ahlobene (ngokuvamile ezikalini zelogi).

Bhala phansi inqubo yakho yokuhlola ukuze abanye bakwazi ukuyiphinda. Ayithandeki—kodwa ikugcina uthembekile.

Yakha vs ukuthenga-nokuthi uzoyisebenzisa kuphi 🏗️

• Ifu : Kulula kakhulu ukuqala, kuhle kakhulu emisebenzini eminingi. Qaphela izindleko zokuphuma.

• Amadivayisi e-Edge : Ukulibaziseka okuphansi kanye nobumfihlo obungcono. Uzokhathalela ukulinganisa, ukuthena, kanye nokusheshisa.

• Iselula ekudivayisi : Iyamangalisa uma ingena. Lungiselela amamodeli bese ubuka ibhethri.

• I-Hybrid : Isihlungi kusengaphambili onqenqemeni, ukuphakamisa okunzima efwini. Ukuvumelana okuhle.

Isitaki esinokwethenjelwa nesidina: i-prototype ene-PyTorch, iqeqesha umshini wokuthola ojwayelekile, ithumele ku-ONNX, isheshise nge-OpenVINO/ONNX Runtime, bese isebenzisa i-OpenCV ukuze icubungulwe kusengaphambili kanye ne-geometry (ukulinganisa, i-homography, i-morphology). [4]

Izingozi, izimiso zokuziphatha, kanye nezingxenye ezinzima ukukhuluma ngazo ⚖️

Izinhlelo zokubona zingazuza ubandlululo lwedatha noma izindawo ezingabonakali zokusebenza. Ukuhlolwa okuzimele (isb., i-NIST FRVT) kulinganise umehluko wabantu emazingeni amaphutha okubona ubuso kuwo wonke ama-algorithms nezimo. Leso akusona isizathu sokwesaba, kodwa kuyisizathu sokuhlola ngokucophelela, ukubhala phansi imikhawulo, nokuqapha njalo ekukhiqizeni. Uma usebenzisa amacala okusetshenziswa ahlobene nobunikazi noma ukuphepha, faka izindlela zokubuyekeza kanye nokudlulisa isikhalazo kubantu. Ubumfihlo, imvume, kanye nokucaca akuzona izinto ezengeziwe ezingakhethwa. [5]

Umhlahlandlela wokuqala okusheshayo ongawulandela ngempela 🗺️

1. Chaza isinqumo
   Yisiphi isinyathelo okufanele uhlelo lusithathe ngemva kokubona isithombe? Lokhu kukuvimbela ekwenzeni ngcono izilinganiso ze-vanity.

2. Qoqa isethi yedatha engabonakali
   Qala ngezithombe ezingamakhulu ambalwa ezibonisa indawo yakho yangempela. Lebula ngokucophelela - noma ngabe nguwe kanye namanothi amathathu anamathelayo.

3. Khetha imodeli eyisisekelo
   Khetha umgogodla olula onezisindo eziqeqeshwe kusengaphambili. Ungalandeli izakhiwo ezingavamile okwamanje. [1]

4. Qeqesha, bhala phansi, hlola
   Landelela amamethrikhi, amaphuzu okudideka, kanye nezindlela zokwehluleka. Gcina incwadi yamanothi “amacala angavamile” - iqhwa, ukukhanya, ukuzindla, amafonti angajwayelekile.

5. Qinisa iluphu
   Engeza ama-hard negatives, lungisa ukugeleza kwelebula, lungisa ama-augmentation, bese ulungisa kabusha imikhawulo. Ukulungiswa okuncane kuyahlangana. [3]

6. Sebenzisa inguqulo encane
   ethi Quantize bese uthumela ngaphandle. Linganisa ukubambezeleka/ukudlula endaweni yangempela, hhayi ibhentshimakhi yethoyizi.

7. Gada futhi ubuyekeze
   Qoqa amaphutha, ulebule kabusha, uqeqeshe kabusha. Hlela ukuhlolwa ngezikhathi ezithile ukuze imodeli yakho ingaguquki.

Icebiso lochwepheshe: chaza isethi encane yokubamba ebekwe yiqembu lakho elingathembisi kakhulu. Uma bengakwazi ukubhoboza, cishe usukulungele.

Izinto ezivamile ongafuna ukuzigwema 🧨

• Ukuqeqeshwa ngezithombe zesitudiyo ezihlanzekile, ukusakaza ezweni langempela elinemvula ilensi.

• Ukuthuthukisa i-mAP iyonke uma ukhathalela ngempela isigaba esisodwa esibalulekile. [3]

• Ukunganaki ukungalingani kwezigaba bese uzibuza ukuthi kungani izehlakalo ezingavamile zinyamalala.

• Ukwandisa kakhulu kuze kube yilapho imodeli ifunda izinto zobuciko zokwenziwa.

• Ukweqa ukulinganiswa kwekhamera bese ulwa namaphutha okubuka unomphela. [4]

• Ukukholelwa ezinombolo zebhodi labaphambili ngaphandle kokuphinda ukusetha okuqondile kokuhlola. [2][3]

Imithombo efanele ukubhukimakhiswa 🔗

Uma uthanda izinto eziyinhloko kanye namanothi ezifundo, lokhu kuyigolide lezisekelo, ukuzijwayeza, kanye nezilinganiso. Bheka isigaba Sezinkomba ukuze uthole izixhumanisi: amanothi e-CS231n, iphepha lenselele le-ImageNet, isethi yedatha/amadokhumenti okuhlola e-COCO, amadokhumenti e-OpenCV, kanye nemibiko ye-NIST FRVT. [1][2][3][4][5]

Amazwi okugcina - noma amade kakhulu, angifundanga 🍃

Umbono Wekhompyutha ku-AI uguqula amaphikseli abe izinqumo. Kukhanya uma uhlanganisa umsebenzi ofanele nedatha efanele, ulinganisa izinto ezifanele, futhi uphindaphinda ngokuziphatha okungavamile. Amathuluzi ayaphana, izilinganiso zisobala, futhi indlela esuka kumodeli kuya ekukhiqizweni imfushane ngokumangazayo uma ugxila esinqumweni sokugcina. Qondanisa amalebula akho, khetha izilinganiso ezihambisana nomthelela, bese uvumela amamodeli ukuthi enze umsebenzi osindayo. Futhi uma isingathekiso sisiza - cabanga ngakho njengokufundisa umfundi oqeqeshwayo osheshayo kodwa ongokoqobo ukubona okubalulekile. Ubonisa izibonelo, ulungisa amaphutha, futhi kancane kancane uthembele emsebenzini wangempela. Akuphelele, kodwa kuseduze ngokwanele ukuba kube noguquko. 🌟

Izinkomba

1. I-CS231n: Ukufunda Okujulile Kombono Wekhompyutha (amanothi ezifundo) - I-Stanford University.
   funda kabanzi

2. Inselele Yokuqashelwa Kokubona Okubonakalayo Okukhulu Ye-ImageNet (iphepha) - uRussakovsky nabanye
   funda kabanzi

3. Isethi Yedatha ye-COCO Nokuhlola - Isayithi elisemthethweni (izincazelo zomsebenzi kanye nemigomo ye-mAP/IoU).
   funda kabanzi

4. Imibhalo ye-OpenCV (v4.x) - Amamojula okucubungula kusengaphambili, ukulinganisa, ukuma, njll.
   funda kabanzi

5. I-NIST FRVT Ingxenye 3: Imiphumela Yezibalo Zabantu (NISTIR 8280) - Ukuhlolwa okuzimele kokunemba kokubonwa kobuso kuzo zonke izinhlaka zabantu.
   funda kabanzi

Thola i-AI Yakamuva Esitolo Esisemthethweni Somsizi we-AI

Mayelana NATHI