Kuonekwa kwechinhu imhando yezvikamu zvemufananidzo umo neural network inotarisira zvinhu mumufananidzo uye inodhirowa mabhokisi anosungirira pavari. Kuona nekuisa zvinhu munzvimbo mumufananidzo unoenderana neyakagadzikwa seti yemakirasi inodaidzwa sekuonekwa kwechinhu.
Kuonekwa kwechinhu (kunozivikanwawo sekuzivikanwa kwechinhu) inonyanya kukosha subdomain yeComputer Vision nekuti mabasa akaita sekuona, kuzivikanwa, uye nharaunda anowana kushanda kwakakura mumamiriro epasirese.
Nzira yeYOLO inogona kukubatsira kuita mabasa aya. Muchinyorwa chino, tichatarisisa paYOLO, kusanganisira kuti chii, kuti inoshanda sei, kusiyana kwakasiyana, nezvimwe.
Saka, chii chinonzi YOLO?
YOLO inzira yekuzivikanwa kwechinhu chaicho-nguva uye kuzivikanwa mumifananidzo. Icho chidimbu chekuti Iwe Chete Tarisa Kamwe. Redmond et al. akakurudzira maitiro mubepa rakatanga kuburitswa muna 2015 paIEEE/CVF Musangano paComputer Vision uye Pattern Recognition (CVPR).
Mubairo weOpenCV People's Choice wakapihwa bepa. Kusiyana nemaitiro ekare ekuzivikanwa kwechinhu, ayo akadzokorora maclassifiers kuti aone, YOLO inokurudzira kushandiswa kwekupedzisira-ku-kuguma. neural network iyo inofanotaura mabhokisi ekusunga uye mikana yekirasi panguva imwe chete.
YOLO inogadzira mhedzisiro yemazuva ano nekutora nzira nyowani yekucherechedzwa kwechinhu, ichipfuura nzira dzekare dzenguva chaiyo dzekuona zvinhu.
YOLO kushanda
Iyo YOLO nzira inokamura mufananidzo kuita N grids, imwe neimwe iine yakaenzana-saizi SxS dimensional chikamu. Imwe neimwe yeaya maN grid ane basa rekuona nekutsvaga chinhu icho chirimo.
Aya magridi, zvakare, anofanotaura B anosungirira bhokisi anoronga zvinoenderana nesero coordinates, pamwe nezita rechinhu uye mukana wechinhu chiripo muchitokisi. Nekuda kwemaseru mazhinji anofanotaura chinhu chimwe chete neakasiyana anosungirirwa mabhokisi ekufanotaura, nzira iyi inodzikisira komputa nekuti zvese kuona uye kuzivikanwa zvinobatwa nemaseru kubva pamufananidzo.
Nekudaro, inoburitsa akawanda ekufanotaura kwakafanana. Kugadzirisa dambudziko iri, YOLO inoshandisa Non-Maximal Suppression. YOLO inodzvanya mabhokisi ese anosungirirwa ane yakaderera mukana zvibodzwa muNon-Maximal Suppression.
YOLO inoita izvi nekuongorora zvingangoitika zvibodzwa zvakabatanidzwa nesarudzo yega yega uye kusarudza ine zvibodzwa zvepamusoro. Mabhokisi anosungirirwa ane Mharadzano huru pamusoro peMubatanidzwa neiyo iripo yepamusoro mukana wekubatanidza bhokisi anobva atsikirirwa.
Iyi nzira inoenderera mberi kusvikira mabhokisi ekusungira apera.
Kusiyana kwakasiyana kweYOLO
Tichatarisa mamwe emhando dzakajairika dzeYOLO. Ngatitangei.
1. YOLOv1
Yekutanga YOLO vhezheni yakaziviswa muna 2015 mukudhindwa "Iwe Unongotarisa Kamwechete: Yakabatana, Chaiyo-Nguva Yekuonekwa Chinhu” naJoseph Redmon, Santosh Divvala, Ross Girshick, naAli Farhadi.
Nekuda kwekumhanya kwayo, huchokwadi, uye kugona kudzidza, YOLO yakakurumidza kutonga nzvimbo yekuzivikanwa kwechinhu uye yakave iyo inonyanya kushandiswa algorithm. Panzvimbo pekutaura nezvekuonekwa kwechinhu senyaya yemhando, vanyori vakaibata sedambudziko rekudzoreredza nemabhokisi akapatsanurwa munzvimbo uye anosanganisirwa kirasi mikana, iyo yavakagadzirisa vachishandisa imwe chete. neural network.
Iyo YOLOv1 yakagadziridzwa mafoto pamafuremu makumi mana neshanu pasekondi mune chaiyo-nguva, nepo diki mutsauko, Fast YOLO, yakagadziriswa pa45 mafuremu pasekondi uye ichiri kuwana yakapetwa kaviri mepu yemamwe-chaiyo-nguva madetector.
2. YOLOv2
Kwapera gore, muna 2016, Joseph Redmon naAli Farhadi vakaburitsa YOLOv2 (inozivikanwawo seYOLO9000) mupepa "YOLO9000: Zvirinani, Zvinokurumidza, Zvakasimba. "
Kugona kwemodhi iyi kufanotaura kunyange zviuru mazana mapfumbamwe emhando dzezvinhu uchiri kushanda munguva-chaiyo zvakauwanira zita rekuti 9000. Haisi chete modhi itsva yakadzidziswa panguva imwe chete pakuona zvinhu uye kurongedza datasets, asi yakawanawo Darknet-9000 seyokutanga. model.
Nekuti YOLOv2 yaivewo budiriro huru uye yakakurumidza kuve inotevera-ye-the-art yekuzivikanwa modhi, mamwe mainjiniya akatanga kuedza nealgorithm uye kugadzira yavo, yakasarudzika YOLO shanduro. Zvimwe zvacho zvichakurukurwa pamapfundo akasiyana mupepa.
3. YOLOv3
Mupepa "YOLOv3: Kuwedzera Kuvandudza, "Joseph Redmon naAli Farhadi vakabudisa shanduro itsva yegorgorithm muna 2018. Yakavakwa paDarknet-53 architecture. Akazvimiririra logistic classifiers akatsiva iyo softmax activation mechanism muYOLOv3.
Iyo binary cross-entropy kurasikirwa yakashandiswa panguva yekudzidziswa. Darknet-19 yakasimudzirwa uye yakatumidzwa zita rekuti Darknet-53, iyo parizvino ine 53 convolutional layer. Kunze kweizvozvo, kufanotaura kwakaitwa pazviyero zvitatu zvakasiyana, izvo zvakabatsira YOLOv3 kusimudzira huroyi hwayo mukufanotaura zvinhu zvidiki.
YOLOv3 yaive yaJoseph Redmon yekupedzisira YOLO vhezheni, sezvo akasarudza kusashanda pane imwezve YOLO kuvandudzwa (kana kunyangwe munzvimbo yekuona komputa) kuitira kuti basa rake rive nekukanganisa nyika. Ikozvino inonyanya kushandiswa senzvimbo yekutanga kuvaka yakasarudzika-yekuona zvivakwa.
4. Yolov4
Alexey Bochkovskiy, Chien-Yao Wang, naHong-Yuan Mark Liao vakaburitswa “YOLOv4: Yakanyanya Kumhanya uye Kurongeka kweKuona Chinhu” muna Kubvumbi 2020, yaive yechina iteration yeYOLO algorithm.
Weighted Residual Connections, Cross-Stage-Partial Connections, cross mini-batch normalization, self-adversarial training, mish activation, drop block, uye CIoU kurasikirwa zvose zvakaunzwa sechikamu cheSDarknet53 architecture.
YOLOv4 muzukuru wemhuri yeYOLO, zvisinei, yakagadzirwa nemasayendisiti akasiyana (kwete Joseph Redmon naAli Farhadi). SPDarknet53 musana, piramidhi yepakati, PANet nzira-kuunganidza semutsipa, uye YOLOv3 musoro unogadzira dhizaini yayo.
Nekuda kweizvozvo, kana ichienzaniswa nemubereki wayo, YOLOv3, YOLOv4 inowana 10% yepamusoro Average Precision uye 12% zvirinani Maframe Per Second metrics.
5. YOLOv5
YOLOv5 ipurojekiti yakavhurika-sosi inosanganisira huwandu hwemamodheru ekuzivikanwa kwechinhu uye algorithms zvichibva paYOLO modhi yakambodzidziswa paCOCO dataset.
YOLOv5 muunganidzwa wemakomboni-scaled zvibodzwa zvemhando yezviro akadzidziswa paCOCO dataset, ine nyore kugona kweTTA, modhi gungano, hyperparameter kuvandudza, uye kutumira kunze kune ONNX, CoreML, uye TFLite. Nekuti YOLOv5 haiite kana kugadzira chero nzira dzakasiyana, bepa repamutemo harina kuburitswa. Ingori YOLOv3's PyTorch yekuwedzera.
Ultranytics yakatora chiitiko ichi kushambadza "itsva YOLO" vhezheni pasi perutsigiro rwayo. Nekuti kune zvakare mamodheru mashanu akafanodzidziswa anowanikwa, iyo YOLOv5 peji rekutanga rakatwasuka uye nehunyanzvi rakarongwa uye rakanyorwa, riine zvidzidzo zvakati wandei nemazano ekudzidzisa nekushandisa mhando dzeYOLOv5.
YOLO zvisingakwanisi
Kunyangwe YOLO ichiita seyakanyanya nzira yekugadzirisa kuona kwechinhu matambudziko, ine chiverengero chezvipingamupinyi. Nekuti grid yega yega inongokwanisa kuona chinhu chimwe chete, YOLO inonetseka kuona uye kupatsanura zvinhu zvidiki mumifananidzo inoitika mumapoka. Zvinhu zvidiki mumapunga, zvakaita semajuru, zvinonetsa kuti YOLO aone uye awane.
Kana ichienzaniswa neinononoka zvakanyanya nzira yekuzivisa chinhu seFast RCNN, YOLO inotaridzwawo nekushomeka kwechokwadi.
Tanga kushandisa YOLOv5
Kana iwe uchida kuona YOLOv5 ichiita, tarisa iyo pamutemo GitHub uye YOLOv5 muPyTorch.
mhedziso
YOLOv5's yekutanga vhezheni inokurumidza zvakanyanya, inoita, uye iri nyore kushandisa. Nepo YOLOv5 isingawedzeri chero mhando nyowani yekuvaka kumhuri yeYOLO, inopa hutsva hwePyTorch kudzidziswa uye dhizaini inosimudzira mamiriro ehunyanzvi hwekuona zvinhu.
Uyezve, YOLOv5 inonyanya kushandisa-mushandisi uye inouya "kunze kwebhokisi" yakagadzirira kushandiswa pazvinhu zvakasarudzika.
Leave a Reply