TeleOCR Uyasebenza

TeleOCR iyimodeli yokuhlaziya amadokhumenti ekhishwe njenge-open source yi-China Telecom Xingchen Laboratory. Isebenzisa ukwakheka kolimi olubonayo olulula olunamapharamitha acishe abe ngu-1.2B, futhi iphatha amadokhumenti edijithali kanye nalawo athathwe ngekhamera ngaphakathi kohlaka olulodwa, ikwazi ukuhumusha umbhalo, ukwakheka kwekhasi, amathebula, amafomula namashadi esayensi kube yimiphumela ehlelekile efana ne-Markdown ne-JSON. I-TeleOCR ihlezi phezulu ohlwini lwe-OmniDocBench v1.6, yanqoba i-PureDocBench kanye nomqhudelwano wokuhlaziya amashadi esayensi we-ICDAR-2026, isekela ukuthunyelwa ku-GPU yendawo futhi ifanele izimo zokukhipha idatha ehlelekile ezinkontilekeni, emaphepheni ocwaningo nasemibhalweni egciniwe.

Ukuqashelwa kombhalo: Ikhipha ngokunemba okuphezulu okuqukethwe kombhalo kumadokhumenti, ngokunemba okungu-97.22 kumadokhumenti edijithali.
Ukuhlaziywa kwesakhiwo sekhasi: Ibeka indawo nobudlelwano bezinga bezinto ezifana nombhalo, amathebula namafomula ekhasini.
Ukuhlaziywa kwamathebula: Ibuyisa ngokunembile izakhiwo zamathebula eziyinkimbinkimbi ezinemigqa namakholomu awelanayo namaseli ahlanganisiwe, i-TEDS ifinyelela ku-97.05.
Ukuhlaziywa kwamafomula: Ihlaziya incazelo nohlelo lwamafomula ngendlela ehlelekile, ibuyise izimpawu ngokunemba kwezinga lezimpawu, i-CDM ifinyelela ku-96.36.
Ukuhlaziywa kwamashadi esayensi: Ihlonza izakhi zamashadi futhi ikhiphe idatha, ingaguqula ishadi lamabha ephepheni libe yithebula elihlelekile, okuyikhono elinqobe i-ICDAR-2026.
Ukucubungula amadokhumenti athathwe ngekhamera: Iphatha ngokuqondile amadokhumenti anokuphazamiseka kwe-perspective, ukugoba kwekhasi, izithunzi nokufiphala, ngaphandle kwemodeli ezimele yokulungisa ukuphazamiseka.
Ukukhishwa okuhlelekile: Imiphumela ingakhiqizwa njenge-Markdown, JSON, amathebula namafomula, okwenza kube lula ukuyixhumanisa nokusesha ulwazi nezinhlelo zebhizinisi.

Landela i-WeChat bese uphendula ngokuthi “开源” ukuze ujoyine iqembu lokushintshisana ngamaphrojekthi e-AI open source
Ukwakhiwa kolimi olubonayo okuhlanganisiwe: I-TeleOCR isebenzisa imodeli yolimi olubonayo olulula olunamapharamitha acishe abe ngu-1.2B, ifake amadokhumenti edijithali nalawo athathwe ngekhamera ohlakeni olulodwa, iqede ukuhlaziya kwesakhiwo nokuqukethwe kusukela ekuqaleni kuya ekugcineni, ithathe indawo yokulandelana kwendabuko ethi “ukutholwa kwesakhiwo→ukulungiswa kwesithombe→ukuqashelwa” futhi igweme ukuqoqeka kwamaphutha.
Ukumodela okwazi ijiyomethri: Ukuze ibhekane nokuguquka kwamakhasi okungajwayelekile, yethula ukumodela okwazi ijiyomethri kanye nokusampula kwe-Douglas-Peucker okuqondiswa yi-curvature, ifunde ngokusobala imingcele yedokhumenti nesakhiwo sendawo, ukuze ikwazi ukuphatha ngokuqondile amakhasi agobile, agoqiwe nanokuphazamiseka kwe-perspective.
Umjikelezo ophelele wobunjiniyela bedatha: Uhlelo lwedatha luhlanganisa ukukhiqiza, ukuhlanza, ukuhlola nokwenza kahle: amavoti okuvumelana ezindaweni eziningi akhetha amalebula-mbumbulu anokuzethemba okuphezulu; ukuhlanganiswa kwedatha okwazi ijiyomethri kwakha amasampula azungezayo, aphazamisekile nagobile; ukuziqinisekisa kwesithombe kuya esithombeni kuphinda kunikeze imiphumela yokuhlaziya ukuze kuqhathaniswe futhi kulungiswe amaphutha ngokuzenzakalelayo; ukuhlanzwa okuqhubekayo kufaka amasampula anokuvumelana okuphezulu kudathasethi yokuqeqesha, kanti amasampula anokwehluka okukhulu asetshenziselwa ukwenza kahle izimo ezinzima, kwakhe umjikelezo wokukhetha ngamavoti—ukuziqeqesha kwemodeli.
Ukuqeqeshwa okuqhubekayo ngezigaba ezine: Ukuqeqeshwa kuqhubeka ngokuthi “ukuhlanganisa umbono nolimi→ukuhlaziya amadokhumenti okuqonda ijiyomethri→ukufunda okuhlukanisa okuqukethwe nesakhiwo→ukufunda okuqinisa”, ngenhloso yokusuka ekuboneni okuyisisekelo kuye ekuqondeni okunemininingwane. Emisebenzini yamathebula, i-Structure-only iqinisa ukuqonda kwesakhiwo semigqa namakholomu, bese i-Structure+Content iqedela ukuvumelanisa okuhlangene; nasemisebenzini yamafomula, ukufunda okuhlukanisiwe kusetshenziswa ukuhlela incazelo nohlelo.

Ukulungisa imvelo: Faka i-Python 3.10+ futhi uqinisekise ukuthi kunemvelo ye-GPU esekela i-CUDA.
Ukulanda ikhodi: Sebenzisa i-git clone https://github.com/caipeng328/TeleOCR.git bese ungene enkombeni.
Ukwakha imvelo: Sebenzisa i-conda ukwakha imvelo ebonakalayo ebizwa nge-teleocr bese uyenza isebenze.
Ukufaka okuncikile: Qalisa i-pip install -e . ukuze ufake okuncikile kwephrojekthi.
Ukulanda imodeli: Ngemva kokufaka i-modelscope nge-pip install modelscope, sebenzisa umyalo we-modelscope ukulanda izisindo uzibeke endaweni yendawo ethi ./models/TeleOCR.
Ukulungisa izinkomba: Setha inkomba yezithombe ezifakwayo IMAGE_SUB_PATH kanye nenkomba yokulondoloza imiphumela RESULT_SAVE_PATH.
Ukuqalisa inference: Qalisa i-python infer.py, udlulise indlela yesithombe, indlela yemiphumela, --use_async, --override, indlela yemodeli kanye nemingcele ye-backend.
Ukukhetha i-backend: Ngokusetshenziswa okuvamile sebenzisa i-vllm-engine; ekucubunguleni okuningi ngesikhathi esisodwa shintshela ku-vllm-async-engine.
Ukusetha imodi yesakhiwo: Kumadokhumenti edijithali sebenzisa LAYOUT_MODE="Detection"; kumakhasi athathwe ngekhamera agobile noma awonakele shintshela ku-"Segmentation".
Ukulungisa imingcele: Lungisa njengoba kudingeka i-MAX_MODEL_LEN, i-GPU_MEMORY_UTILIZATION, i-PDF_TOOLS kanye ne-MAX_PIXELS.
Ukuthola imiphumela: Funda emiphumeleni ehlelekile ye-Markdown/JSON/amathebula nokunye ekwifolda yokukhipha, bese uyixhumanisa nezinhlelo ezilandelayo.

Imiphumela ehamba phambili: Ithole amaphuzu angu-96.87 ku-OmniDocBench v1.6 yaphinde yaba phezulu, yedlula i-MinerU2.5-Pro, i-PaddleOCR-VL-1.6, i-OvisOCR2 namanye amamodeli ochwepheshe.
Inamapharamitha ambalwa: Inamapharamitha acishe abe ngu-1.2B kuphela, yehlisa ubunzima bokuthunyelwa kwendawo futhi ifanele izimo ezizimele.
Uhlaka oluhlanganisiwe: Amadokhumenti edijithali nalawo athathwe ngekhamera ahlaziywa kumodeli efanayo, ngaphandle kwesidingo sokuthunyelwa kwezinhlelo eziningi.
Ayidingi ukulungiswa kwangaphambili: Ukumodela okwazi ijiyomethri kuphatha ngokuqondile amakhasi agobile, agoqiwe nanokuphazamiseka kwe-perspective, kunciphise ukuqoqeka kwamaphutha kumamojuli azimele okulungisa.
Ukukhishwa okuhlelekile: Ikhipha ngokuqondile imiphumela ebalekayo efana ne-Markdown, JSON, amathebula namafomula, okwenza kube lula ukuxhuma ekusesheni ulwazi nasekuzenzakaleleni kwebhizinisi.
Ikhono eliqinile kumathebula ayinkimbinkimbi: Isakhiwo semigqa namakholomu samathebula anemigqa namakholomu awelanayo namaseli ahlanganisiwe sihlala silungile futhi singaxubani, i-TEDS ifinyelela ku-97.05.
Inqobe ekuhlaziyweni kwamashadi esayensi: Iyakwazi ukuhlonza izakhi zamashadi nokukhipha idatha, yanqoba umqhudelwano wokuhlaziya amashadi esayensi we-ICDAR-2026.
Umjikelezo wedatha osebenza kahle: Amavoti okuvumelana ezindaweni eziningi kanye nendlela yokuziqinisekisa kwakha ichibi ledatha yokuqeqesha elinamashumi ezigidi zamasampula, futhi iningi ledatha ekhiqiziwe alidingi ukuphawulwa ngesandla.

Inqolobane ye-GitHub: https://github.com/caipeng328/TeleOCR
Inqolobane yamamodeli ye-HuggingFace: https://www.modelscope.cn/models/XingChen-AGI/TeleOCR
Iphepha lobuchwepheshe le-arXiv: https://arxiv.org/pdf/2608.12898

Ukwenziwa kwedijithali kwezinkontileka nemibhalo egciniwe: Ihlaziya amathebula anemigqa namakholomu awelanayo kanye nesakhiwo samakhasi ezinkontilekeni nasemibhalweni egciniwe, ikuguqulele kudatha ehlelekile eseshisekayo.
Ukuhlaziywa kwamaphepha ocwaningo: Ikhipha amafomula, amathebula nedatha yamashadi amabha emaphepheni ocwaningo, isize ekuhlaziyweni kwemibhalo nasekwakheni isizindalwazi solwazi.
Ukucubungula amaphepha ezezimali: Ihlonza izithombe zama-invoyisi namafomu, ibekezelele ukuphazamiseka kwe-perspective nokufiphala, bese ifaka idatha ngokuzenzakalelayo ohlelweni lwebhizinisi.
Ukufakwa kwemibuzo yezemfundo: Iguqula amafomula aphrintiwe namathebula ezivivinyweni abe amafomethi ahlelekayo, isekele ukwakhiwa kwamabhange emibuzo nokumaka okuhlakaniphile.
Ukuzenzakalela kwemibiko yezinkampani: Iguqula ngobuningi imibiko eskeniwe namafomu ebhizinisi ibe amathebula ahlelekile, iwaxhumanise nokuhlaziywa kwedatha nezinqubo ze-RPA.