Niba wigeze gufungura telefoni yawe ukoresheje isura yawe, ugasuzuma inyemezabwishyu, cyangwa ukitegereza kamera yiyishyurira wibaza niba ari ugusuzuma avoka yawe, wagerageje kureba mudasobwa. Muri make, Computer Vision muri AI ni uburyo imashini ziga kubona no gusobanukirwa amashusho na videwo neza bihagije kugira ngo zifate ibyemezo. Ni ingirakamaro? Nibyo rwose. Rimwe na rimwe biratangaje? Yego kandi. Kandi rimwe na rimwe biratangaje gato niba tuvugisha ukuri. Mu buryo bwiza, ihindura pixels zitameze neza mo ibikorwa bifatika. Mu buryo bubi cyane, iratekereza kandi igahindagurika. Reka dukore ubushakashatsi neza.
Ingingo ushobora gukunda gusoma nyuma y'iyi:
🔗 Ni iki gitandukanya ikoranabuhanga rya AI?
Uburyo bias igaragara muri sisitemu za AI n'uburyo bwo kuyimenya no kuyigabanya.
🔗 Uburyo
bwo gukora isuzuma ry’impinduka mu buryo bw ...
🔗 Ni iki umwigisha w’ubuhanga
mu by ...
🔗 ya Google Vertex AI
ku rubuga rwa Google ruhuriweho rwa AI rwo kubaka no gukoresha moderi.
Computer Vision muri AI ni iki, mu by'ukuri? 📸
Iyerekwa rya mudasobwa muri AI ni ishami ry'ubwenge bw'ubukorano ryigisha mudasobwa gusobanura no gutekereza ku makuru aboneka. Ni inzira kuva kuri pikseli zitari umwimerere kugeza ku nsobanuro yubatswe: "iki ni ikimenyetso cyo guhagarara," "abo ni abanyamaguru," "gusudira bifite inenge," "igiteranyo cy'inyemezabwishyu kirahari." Ikubiyemo imirimo nko gushyira mu byiciro, gutahura, gutandukanya, gukurikirana, gupima uburebure, OCR, n'ibindi bivanze hamwe hakoreshejwe moderi zo kwiga ibishushanyo. Urwego rwemewe rugizwe n'imiterere ya kera kugeza ku bumenyi bugezweho, hamwe n'ibitabo by'imikino ushobora gukoporora no kubikosora. [1]
Inkuru yihuse: tekereza umurongo wo gupakira ufite kamera ntoya ya 720p. Akamashini gapima ibintu byoroheje gafata imigozi, maze agakoresho koroshye ko gukurikirana ibintu gakemeza ko byashyizwe ku murongo mu mashusho atanu yikurikiranya mbere yo gucana icupa. Ntabwo ari ibintu bigezweho - ariko bihendutse, ni byihuse, kandi bigabanya gusubiramo.
Ni iki gituma Computer Vision muri AI iba ingirakamaro? ✅
-
Uburyo bwo gukoresha ikimenyetso mu gukora : Ishusho yinjiye iba umusaruro ushobora gukorwa. Dashboard nke, ni bwo hafatwa ibyemezo byinshi.
-
Guhuza amakuru : Iyo habayeho amakuru akwiye, icyerekezo kimwe gikemura ibibazo bitandukanye by’amashusho. Ntabwo ari byiza cyane - rimwe na rimwe bitangaje.
-
Ingufu z'amakuru : Kamera zirahendutse kandi ziboneka hose. Iyerekwa rihindura iyo nyanja ya pikseli mo ubumenyi.
-
Umuvuduko : Moderi zishobora gutunganya amafuremu mu gihe nyacyo ku bikoresho bito cyangwa hafi yabyo, bitewe n'akazi n'uburyo byakemutse.
-
Uburyo bwo guhuza ibintu : Guhuza intambwe zoroshye muri sisitemu zizewe: kumenya → gukurikirana → kugenzura ubuziranenge.
-
Urusobe rw'ibidukikije : Ibikoresho, ingero zateguwe mbere, ibipimo ngenderwaho, n'ubufasha mu baturage - isoko rinini ry'amategeko.
Tuvugishe ukuri, ibanga si ibanga: amakuru meza, isuzumabumenyi rihamye, gushyira mu bikorwa witonze. Ibisigaye ni imyitozo ... wenda ikawa. ☕
Uburyo Computer Vision muri AI ikora, mu buryo bumwe buboneye 🧪
-
Gufata amashusho
Kamera, scanner, drones, telefoni. Hitamo ubwoko bwa sensor, uko igaragara, lens, n'umuvuduko wa frame witonze. Shyira imyanda muri byo, nibindi. -
Gutunganya mbere y’uko
ibintu bihinduka, gukata, gutunganya, gukuraho urusaku cyangwa gukuraho urusaku nibiba ngombwa. Hari igihe guhindura gato itandukaniro ry’imiterere y’ibintu bihindura imisozi. [4] -
Ibirango n'amakuru
Udusanduku dufatanye, poligoni, ingingo z'ingenzi, imirongo. Ibirango bingana kandi byerekana neza - cyangwa icyitegererezo cyawe cyiga ingeso zidafite aho zihuriye. -
Gukora ibishushanyo mbonera
-
Ishyirwa mu byiciro : “Ni ikihe cyiciro?”
-
Gutahura : "Ibintu biri he?"
-
Gutandukanya : “Ni izihe pixels zigize iki kintu?”
-
Ingingo z'ingenzi n'imiterere : “Aho hantu cyangwa ahantu nyaburanga biri?”
-
OCR : “Ni iyihe nyandiko iri ku ishusho?”
-
Ubujyakuzimu na 3D : "Ikintu cyose kiri kure gute?"
Imiterere y'inyubako iratandukanye, ariko imbuga za convolutional na models zo mu bwoko bwa transformer ziganza. [1]
-
-
Guhugura
Gutandukanya amakuru, gutunganya ibipimo bya hyperparameter, gutunganya no kongera. Guhagarara kare mbere yo gufata mu mutwe urupapuro rwo gushushanya. -
Isuzuma
Koresha ibipimo bijyanye n'inshingano nka mAP, IoU, F1, CER/WER kuri OCR. Ntugahitemo neza. Gereranya neza. [3] -
Gushyira mu bikorwa
Optimize ku ntego: imirimo ya cloud batch, inference kuri tool, edge servers. Genzura drift. Ongera witoze igihe isi ihinduka.
Imbuga zirambuye zateye intambwe y’ingenzi nyuma y’uko amakuru manini n’imibare bigeze ku bwinshi. Ibimenyetso nk’ikibazo cya ImageNet byatumye iyo ntambwe igaragara kandi idashira. [2]
Imirimo y'ingenzi uzakoresha (n'igihe uzayikorera) 🧩
-
Ishyirwa mu byiciro ry'ishusho : Ikirango kimwe kuri buri shusho. Gikoreshwa mu kuyungurura byihuse, kugabanya ibiciro, cyangwa ku marembo meza.
-
Gutahura ibintu : Udusanduku dukikije ibintu. Gukumira igihombo mu maduka, kubona ibinyabiziga, umubare w'inyamaswa zo mu gasozi.
-
Gutandukanya ibintu : Imiterere y'ibishushanyo mbonera by'amashusho kuri buri kintu. Inenge mu gukora, ibikoresho byo kubaga, agritech.
-
Gutandukanya ibice : Icyiciro kuri pikseli nta ngero zitandukanya. Imiterere y'umuhanda wo mu mijyi, igisenge cy'ubutaka.
-
Gutahura ingingo z'ingenzi n'imiterere yazo : Ingingo z'ingenzi, ibimenyetso by'ingenzi, imiterere y'isura. Isesengura rya siporo, imiterere y'umubiri, AR.
-
Gukurikirana : Gukurikirana ibintu uko igihe kigenda. Ibikoresho, urujya n'uruza rw'abantu, umutekano.
-
OCR & AI y'inyandiko : Gukuramo inyandiko no gusesengura imiterere yayo. Inyemezabuguzi, inyemezabwishyu, amafishi.
-
Ubujyakuzimu na 3D : Ivugururwa rishingiye ku buryo butandukanye cyangwa ibimenyetso byerekana ishusho imwe. Robotics, AR, ikarita.
-
Incamake y'amashusho : Shyira muri make amashusho mu mvugo karemano. Uburyo bwo kuyabona, gushakisha.
-
Ingero z'ururimi rw'icyerekezo : Gutekereza mu buryo bwinshi, kureba kwiyongereyeho, QA ishingiyeho.
Akantu gato k'inyuma: mu maduka, agakoresho gapima ibintu gatuma ibintu bitagira aho bihurira n'amabati; agakoresho gapima ibintu gatuma abakozi babarwa kabiri mu gihe cyo kongera gushyiramo ibikoresho; itegeko ryoroshye rituma abantu batizeye neza. Ni itsinda rito ry'abaririmbyi rihora rikora neza.
Imbonerahamwe yo kugereranya: ibikoresho byo kohereza vuba 🧰
Biteye ishozi gato nkana. Yego, intera iri hagati y'ibintu ni idasanzwe - ndabizi.
| Igikoresho / Umushinga | Ibyiza kuri | Uruhushya/Igiciro | Impamvu ikora mu buryo bufatika |
|---|---|---|---|
| OpenCV | Gutegura mbere y'igihe, CV isanzwe, POC zihuse | Ubuntu - isoko ifunguye | Agasanduku k'ibikoresho kanini, API zihamye, zageragejwe n'intambara; rimwe na rimwe ni byo ukeneye gusa. [4] |
| PyTorch | Amahugurwa yoroshye gukora ubushakashatsi | Ubuntu | Imbonerahamwe zihindagurika, imiterere y'ibinyabuzima nini, inyigisho nyinshi. |
| TensorFlow/Keras | Umusaruro ku rwego rwo hejuru | Ubuntu | Uburyo bwo gutanga serivisi zikuze, ni bwiza no ku zigendanwa ndetse no ku zindi. |
| Ultralytics YOLO | Gutahura ibintu byihuse | Inyongera z'ubuntu + zishyurwa | Uburyo bworoshye bwo gutoza, umuvuduko ukwiye mu irushanwa, ufite ibitekerezo bihamye ariko woroshye. |
| Detectron2 / MMDetection | Imiterere ikomeye, igabanywa ry'ibice | Ubuntu | Ingero z'icyitegererezo zifite ibisubizo bishobora kongera gutangwa. |
| OpenVINO / ONNX Runtime | Guteza imbere igenamigambi | Ubuntu | Kanda gutinda, shyira ahagaragara cyane udakoresheje ubundi buryo bwo kwandika. |
| Tesseract | OCR ku ngengo y'imari | Ubuntu | Bikora neza iyo usukuye ishusho… rimwe na rimwe ni byo koko wagombye gukora. |
Ni iki gituma habaho ireme muri Computer Vision muri AI 🔧
-
Gukwirakwiza amakuru : Impinduka z'amatara, inguni, inyuma, impande z'amatara. Niba bishoboka, shyiramo.
-
Ubwiza bw'ikirango : Udusanduku tudahuye cyangwa polygons zitameze neza zirasenya mAP. QA ntoya igira akamaro kanini.
-
Kongera ikoranabuhanga : Kuraho, zungurutsa, shyira urumuri, ongeramo urusaku rw'ubukorano. Ba umuntu ufatika, ntugire akajagari gatunguranye.
-
Uburyo bwo guhitamo icyitegererezo : Koresha uburyo bwo gutahura aho bikenewe ko umenya - ntugahatire umuhanga mu gusesengura kugira ngo amenye aho aherereye.
-
Ibipimo bihuye n'ingaruka : Niba amakosa atari yo yababaje cyane, ongera ubushobozi bwo kwibuka. Niba amakosa atari yo yababaje cyane, banza umenye neza.
-
Uburyo bwo gusubiza ibintu mu buryo buhamye : Kwangirika kw'inyandiko, kongera kwandikaho, kongera gutoza. Koza, ongeramo. Birarambiranye gato-bifite akamaro gakomeye.
Ku bijyanye no kumenya/gutandukanya, igipimo ngenderwaho cy’umuryango ni impuzandengo y’ubuziranenge ikoreshwa mu mibare ya IoU-aka COCO-style mAP . Kumenya uburyo IoU na AP@{0.5:0.95} bibarwa bituma ibirego by’ubuyobozi bitagutangaza cyane bitewe n’imibare y’ibice. [3]
Imikoreshereze y'ibintu mu buryo bufatika atari iyo mu buryo bw'imitekerereze 🌍
-
Ubucuruzi : Gusesengura ububiko, gukumira igihombo, gukurikirana umurongo, kubahiriza planogram.
-
Gukora : Gutahura inenge zo hejuru, kugenzura imiterere y'ibikoresho, kuyobora robo.
-
Ubuvuzi : Gupima ibice by'uturemangingo hakoreshejwe radiyo, kumenya ibikoresho, no gutandukanya uturemangingo.
-
Uburyo bwo kugenda : ADAS, kamera zo mu muhanda, aho imodoka zihagarara, gukurikirana uburyo bwo kugenda mu buryo bwa micromobility.
-
Ubuhinzi : Kubara ibihingwa, kumenya indwara, kwitegura gusarura.
-
Ubwishingizi n'Imari : Isuzuma ry'ibyangiritse, igenzura rya KYC, ibimenyetso by'uburiganya.
-
Ubwubatsi n'Ingufu : Gukurikiza umutekano, kumenya aho amazi aturuka, kugenzura ingese.
-
Ibikubiye muri iyi porogaramu no kubigeraho : Inyandiko zisobanura ibintu mu buryo bwikora, kugenzura, gushakisha mu buryo bw'amashusho.
Imiterere uzayibona: simbura gupima intoki ukoresheje uburyo bwo gukora isuzuma ryikora, hanyuma bizamuka bikagera ku bantu iyo icyizere kigabanutse. Ntabwo ari byiza cyane - ariko birarushaho kwiyongera.
Amakuru, ibirango, n'ibipimo by'ingenzi 📊
-
Ishyirwa mu byiciro : Ubunyangamugayo, F1 ku bijyanye n'ubusumbane.
-
Gutahura : mAP ku bipimo bya IoU; kugenzura buri cyiciro cya AP n'ingano yabyo. [3]
-
Gutandukanya ibice : mIoU, Dice; reba n'amakosa yo mu rwego rwa instance-level.
-
Gukurikirana : MOTA, IDF1; ireme ryo kwigaragaza ni intwari itagira icyo ivuga.
-
OCR : Igipimo cy'ikosa ry'inyuguti (CER) n'igipimo cy'ikosa ry'amagambo (WER); gutsindwa kw'imiterere akenshi ni byo byiganje.
-
Imirimo yo gusubira inyuma : Ubujyakuzimu cyangwa imiterere ikoresha amakosa ashingiye ku buryo busesuye/bufitanye isano (akenshi ku bipimo bya log).
Andika amabwiriza yawe yo gusuzuma kugira ngo abandi bayakore. Ntabwo ari byiza cyane - ariko agufasha kugumana ukuri.
Kubaka vs kugura-n'aho kubikoresha 🏗️
-
Cloud : Yoroshye gutangira, ni nziza ku mirimo myinshi. Reba ikiguzi cyo gusohoka.
-
Ibikoresho byo mu mpande : Kugabanya igihe cyo gutinda no kugira ubuzima bwite bwiza. Uzita ku bijyanye no gupima, gukata, no kwihutisha ibintu.
-
Telefoni igendanwa iri kuri telefoni : Iratangaje iyo ikwiranye. Hindura moderi kandi urebe bateri.
-
Hybrid : Gushungura mbere y'uko ugera ku nkengero, guterura ibintu biremereye mu gicu. Ni uburyo bwiza bwo guhuza ibintu.
Itsinda ryizewe kandi ritarambiranye: prototype ikoresheje PyTorch, itoza icyuma gipima ubuziranenge, yohereza muri ONNX, yihutishe na OpenVINO/ONNX Runtime, kandi ikoreshe OpenCV mu gutunganya no gupima imiterere y'ishusho (gupima, guhuza, imiterere). [4]
Ingaruka, amahame mbwirizamuco, n'ibice bigoye kuvugaho ⚖️
Sisitemu zo kureba zishobora kuragwa ivangura rishingiye ku makuru cyangwa ahantu hatagaragara. Isuzuma ryigenga (urugero, NIST FRVT) ryapimye itandukaniro ry’abaturage mu gipimo cy’amakosa yo kumenya isura mu buryo butandukanye n’imiterere. Iyo si impamvu yo gutinya, ahubwo ni impamvu yo gusuzuma witonze, kwandika aho inzitizi zigarukira, no gukurikirana buri gihe mu ikorwa. Niba ukoresha imanza zijyanye n’irangamuntu cyangwa umutekano, shyiramo uburyo bwo gusuzuma no kujuririra abantu. Ubuzima bwite, ubwumvikane, no gukorera mu mucyo si ibintu by’inyongera. [5]
Gahunda y'intangiriro yihuse ushobora gukurikiza 🗺️
-
Sobanura icyemezo
Ni ikihe gikorwa sisitemu igomba gukora nyuma yo kubona ishusho? Ibi bikurinda kunoza ibipimo by'ubuziranenge. -
Kusanya amakuru asanzwe
Tangira ukoresheje amashusho amagana agaragaza aho utuye nyaho. Shyiraho amazina witonze - nubwo yaba ari wowe n'inyandiko eshatu zifata. -
Hitamo icyitegererezo cy'ibanze
Hitamo umugongo woroshye ufite ibiremereye byatojwe mbere. Ntugakurikirane imiterere idasanzwe. [1] -
Toza, andika, suzuma Kurikirana
ibipimo, aho urujijo rubera, n'uburyo bwo gutsindwa. Bika ikaye y' "ibintu bidasanzwe" - urubura, ubwiza bw'urumuri, ibitekerezo, inyuguti zidasanzwe. -
Kangura umurongo
Ongeramo ibirangaza bikomeye, kosora ihindagurika ry'ibirango, hindura uburyo bwo kongeramo, kandi uhindure urwego rw'ibiciro. Udushya duto twongeraho. [3] -
Shyiraho verisiyo ntoya
. Gupima igihe cyo gutinda/gusohora ibintu mu buryo nyabwo, ntabwo ari igipimo cy'igikinisho. -
Genzura kandi usubiremo
Gukusanya amakosa, ongera ushyireho amazina, ongera utoze. Teganya isuzuma buri gihe kugira ngo icyitegererezo cyawe kidahinduka.
Inama y'inzobere: garagaza uburyo mugenzi wawe ukunda cyane yakwitwara nabi. Niba badashobora kubitobora, ushobora kuba witeguye.
Ibintu bisanzwe ugomba kwirinda 🧨
-
Amahugurwa ku mashusho meza ya studio, gukoresha mu isi nyayo imvura iri ku jisho.
-
Gukoresha mAP muri rusange iyo wita cyane ku cyiciro kimwe cy’ingenzi. [3]
-
Kwirengagiza ubusumbane mu nzego z'abantu hanyuma ukibaza impamvu ibintu bidasanzwe bishira.
-
Kongera imbaraga kugeza igihe icyitegererezo kimenye ibihangano by'ubukorikori.
-
Gusimbuka gupima kamera hanyuma ugakomeza kurwanya amakosa yo kureba ibintu uko biri. [4]
-
Kwizera imibare y’abayobozi badakoresheje uburyo bwo gusuzuma neza. [2][3]
Amasoko akwiye gushyirwaho 🔗
Niba ukunda ibikoresho by'ibanze n'inyandiko z'amasomo, ibi ni zahabu ku by'ibanze, imyitozo, n'ibipimo ngenderwaho. Reba igice ajyanye n'amasomo kugira ngo ubone amasano: inyandiko za CS231n, inyandiko y'ikibazo cya ImageNet, inyandiko z'amakuru/isuzuma rya COCO, inyandiko za OpenCV, na raporo za NIST FRVT. [1][2][3][4][5]
Amagambo ya nyuma - cyangwa igihe kirekire cyane, ntasomye 🍃
Iyerekwa rya mudasobwa muri AI rihindura pixels mu byemezo. Rirabagirana iyo uhuza igikorwa gikwiye n'amakuru akwiye, ugapima ibintu bikwiye, kandi ugasubiramo ibintu bidasanzwe. Uburyo bwo gukoresha ibikoresho ni bwinshi, ibipimo ni ibya rubanda, kandi inzira kuva ku rugero rumwe kugera ku musaruro ni ngufi cyane iyo wibanda ku cyemezo cya nyuma. Hitamo ibirango byawe neza, hitamo ibipimo bihuye n'ingaruka, kandi ureke ingero zikore imirimo iremereye. Kandi niba ikigereranyo gifasha - tekereza nko kwigisha umutoza wihuta cyane ariko usanzwe kumenya icy'ingenzi. Ugaragaza ingero, ugakosora amakosa, kandi buhoro buhoro ukayizera akazi nyako. Ntabwo ari nziza, ariko hafi bihagije kugira ngo ube uhindura ibintu. 🌟
Amareferensi
-
CS231n: Kwiga mu buryo bwimbitse kugira ngo ubone icyerekezo cya mudasobwa (inyandiko z'amasomo) - Kaminuza ya Stanford.
soma byinshi -
ImageNet Large Scale Visual Recognition Challenge (inyandiko) - Russakovsky n'abandi
soma byinshi -
Amakuru ya COCO n'Isuzuma - Urubuga rwemewe (ibisobanuro by'imirimo n'amahame ya mAP/IoU).
soma byinshi -
Inyandiko za OpenCV (v4.x) - Modules zo gutunganya, gupima, imiterere, nibindi.
soma byinshi -
Igice cya 3 cya NIST FRVT: Ingaruka z'abaturage (NISTIR 8280) - Isuzuma ryigenga ry'ubuziranenge bwo kumenya isura mu mibare y'abaturage.
Soma byinshi