{"id":128951,"date":"2025-10-21T11:00:19","date_gmt":"2025-10-21T11:00:19","guid":{"rendered":"https:\/\/kamucalisani.net\/?p=128951"},"modified":"2025-10-21T11:00:19","modified_gmt":"2025-10-21T11:00:19","slug":"deepseek-ocr-yapay-zeka-artik-metinleri-okumuyor-goruyor","status":"publish","type":"post","link":"https:\/\/kamucalisani.net\/index.php\/2025\/10\/21\/deepseek-ocr-yapay-zeka-artik-metinleri-okumuyor-goruyor\/","title":{"rendered":"Deepseek OCR: Yapay zeka art\u0131k metinleri okumuyor, \u201cg\u00f6r\u00fcyor\u201d"},"content":{"rendered":"

<\/span> \u00c7inli yapay zeka \u015firketi Deepseek<\/strong>, metin tabanl\u0131 belgeleri daha verimli i\u015fleyebilmek i\u00e7in geli\u015ftirdi\u011fi yeni OCR<\/strong> (Optik Karakter Tan\u0131ma) sistemiyle dikkat \u00e7ekiyor. Sistem, g\u00f6rsel tabanl\u0131 metinleri s\u0131k\u0131\u015ft\u0131rarak yapay zeka modellerinin \u00e7ok daha uzun belgeleri bellek s\u0131n\u0131r\u0131na tak\u0131lmadan i\u015flemesini m\u00fcmk\u00fcn k\u0131l\u0131yor. <\/figure>\n

Deepseek\u2019in teknik raporuna g\u00f6re bu sistem, metin verilerini do\u011frudan i\u015flemek yerine g\u00f6r\u00fcnt\u00fc bi\u00e7iminde analiz ediyor. Bu yakla\u015f\u0131m, i\u015flem y\u00fck\u00fcn\u00fc \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131yor. Yeni OCR sistemi, metinleri 10 kata kadar s\u0131k\u0131\u015ft\u0131r\u0131rken bilgilerin y\u00fczde 97\u2019sini koruyabiliyor<\/strong>.<\/p>\n

Bilindi\u011fi \u00fczere b\u00fcy\u00fck dil modelleri metni token\u2019lar halinde temsil eder, her token birka\u00e7 karakter i\u00e7erir. K\u0131sa kelimeler genellikle tek bir token ile g\u00f6sterilir (\u201cthe\u201d, \u201cit\u201d), uzun kelimeler ise birden fazla token\u2019a b\u00f6l\u00fcnebilir (\u201cindivisible\u201d \u2192 \u201cind\u201d, \u201civ\u201d, \u201cisible\u201d). Ara\u015ft\u0131rmac\u0131lar milyonlarca token\u2019\u0131 a\u015fan uzun belgeleri ve konu\u015fmalar\u0131 i\u015fleyebilen modeller geli\u015ftirmek i\u00e7in \u00e7al\u0131\u015f\u0131yor. Bu sayede ba\u011flam penceresini geni\u015f tutabiliyorlar. Ancak ayn\u0131 anda i\u015flenebilen token say\u0131s\u0131 artt\u0131k\u00e7a hesaplama maliyetleri de ona g\u00f6re art\u0131\u015f g\u00f6steriyor. Dolay\u0131s\u0131yla b\u00fcy\u00fck token kapasitesi uzun belgelerde bile modelin belle\u011finin dolmamas\u0131n\u0131 sa\u011flar ama maliyeti art\u0131r\u0131r. Deepseek\u2019in OCR \u00e7\u00f6z\u00fcm\u00fc ise \u00e7ok uzun i\u00e7erikleri bir g\u00f6rselmi\u015f gibi i\u015fliyor. Bu sayede i\u00e7erikler esas\u0131nda pikseller olarak g\u00f6r\u00fclm\u00fc\u015f oluyor.<\/p>\n

Uzun yaz\u0131lar\u0131 pikselleri haline g\u00f6r\u00fcyor<\/b><\/p>\n

<\/span> Sistemin \u00e7ekirde\u011finde iki temel bile\u015fen bulunuyor: DeepEncoder<\/strong> ve Deepseek3B-MoE<\/strong>. G\u00f6r\u00fcnt\u00fc i\u015fleme k\u0131sm\u0131n\u0131 \u00fcstlenen DeepEncoder, 380 milyon parametreyle \u00e7al\u0131\u015f\u0131yor. Metin \u00fcretiminden sorumlu Deepseek3B-MoE ise 570 milyon aktif parametreye sahip. DeepEncoder, Meta\u2019n\u0131n 80 milyon parametreli SAM (Segment Anything Model) sistemini ve OpenAI\u2019\u0131n 300 milyon parametreli CLIP modelini birle\u015ftiriyor. Arada yer alan 16x s\u0131k\u0131\u015ft\u0131r\u0131c\u0131, g\u00f6r\u00fcnt\u00fc verilerini b\u00fcy\u00fck \u00f6l\u00e7\u00fcde azaltarak i\u015flem h\u0131z\u0131n\u0131 art\u0131r\u0131yor. \u00d6rne\u011fin, 1.024 x 1.024 piksel boyutundaki bir g\u00f6rselin 4.096 token\u2019\u0131, s\u0131k\u0131\u015ft\u0131rma sonras\u0131nda yaln\u0131zca 256 token\u2019a indiriliyor. <\/figure>\n
<\/span> Deepseek OCR, \u00e7\u00f6z\u00fcn\u00fcrl\u00fc\u011fe g\u00f6re 64 ile 400 aras\u0131nda \u201cvision token\u201d kullanarak \u00e7al\u0131\u015fabiliyor. Bu say\u0131, klasik OCR sistemlerinde genellikle binlerce token gerektiren i\u015flemleri olduk\u00e7a hafif hale getiriyor. OmniDocBench testlerinde sistem, yaln\u0131zca 100 vision token kullanarak GOT-OCR 2.0\u2019\u0131 geride b\u0131rakt\u0131. Ayr\u0131ca 800 token\u2019\u0131n alt\u0131nda \u00e7al\u0131\u015f\u0131rken, MinerU 2.0\u2019\u0131n 6.000\u2019den fazla token gerektiren performans\u0131n\u0131 da a\u015ft\u0131. <\/figure>\n
<\/span> Farkl\u0131 belge t\u00fcrlerine g\u00f6re optimize edilen sistem, basit sunumlarda 64 token, kitap ve raporlarda 100 token, karma\u015f\u0131k gazetelerde ise \u201cGundam modu\u201d ad\u0131 verilen \u00f6zel modla 800 token kullan\u0131yor. Deepseek OCR, yaln\u0131zca metinleri de\u011fil, diyagramlar, kimyasal form\u00fcller ve geometrik \u015fekiller gibi karma\u015f\u0131k g\u00f6rsel unsurlar\u0131 da i\u015fleyebiliyor. Ayr\u0131ca yakla\u015f\u0131k 100 dilde \u00e7al\u0131\u015f\u0131yor, bi\u00e7imlendirmeyi koruyabiliyor ve istenirse d\u00fcz metin ya da genel g\u00f6rsel a\u00e7\u0131klamas\u0131 \u00fcretebiliyor. <\/figure>\n

G\u00fcnde 33 milyon sayfa i\u015fliyor<\/b><\/p>\n

<\/span> Sistemin e\u011fitimi i\u00e7in yakla\u015f\u0131k 30 milyon PDF sayfas\u0131 kullan\u0131ld\u0131. Bu verilerin 25 milyonu \u0130ngilizce ve \u00c7ince belgelerden, geri kalan\u0131 ise 10 milyon sentetik diyagram, 5 milyon kimyasal form\u00fcl ve 1 milyon geometrik \u015fekilden olu\u015fuyor. <\/figure>\n

Ger\u00e7ek d\u00fcnya kullan\u0131m\u0131nda ise Deepseek OCR olduk\u00e7a y\u00fcksek bir i\u015flem kapasitesine ula\u015f\u0131yor. Sistem, tek bir Nvidia A100 GPU \u00fczerinde g\u00fcnde 200.000 sayfadan fazla belgeyi i\u015fleyebiliyor. 20 sunucuda, her biri sekiz A100 GPU bar\u0131nd\u0131rd\u0131\u011f\u0131nda bu kapasite g\u00fcnde 33 milyon sayfaya \u00e7\u0131k\u0131yor. Bu h\u0131z, yeni yapay zeka modelleri i\u00e7in e\u011fitim verisi \u00fcretimini b\u00fcy\u00fck \u00f6l\u00e7\u00fcde kolayla\u015ft\u0131rma potansiyeli ta\u015f\u0131yor. Modelin hem kod hem de model a\u011f\u0131rl\u0131klar\u0131 halka a\u00e7\u0131k durumda. Kaynak\u00e7a k\u0131sm\u0131ndan eri\u015febilirsiniz.<\/p>\n\n

Kaynak\u00a0 :\u00a0https:\/\/www.donanimhaber.com\/deepseek-ocr-yapay-zeka-artik-metinleri-okumuyor-goruyor–197553<\/span><\/p>\n","protected":false},"excerpt":{"rendered":"

\u00c7inli yapay zeka \u015firketi Deepseek, metin tabanl\u0131 belgeleri daha verimli i\u015fleyebilmek i\u00e7in geli\u015ftirdi\u011fi yeni OCR (Optik Karakter Tan\u0131ma) sistemiyle dikkat \u00e7ekiyor. Sistem, g\u00f6rsel tabanl\u0131 metinleri s\u0131k\u0131\u015ft\u0131rarak yapay zeka modellerinin \u00e7ok daha uzun …<\/p>\n","protected":false},"author":1,"featured_media":128952,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[8],"tags":[1118,8597,47,449,6440],"class_list":["post-128951","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-teknoloji","tag-belge","tag-deepseek","tag-milyon","tag-model","tag-token"],"_links":{"self":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts\/128951","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/comments?post=128951"}],"version-history":[{"count":1,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts\/128951\/revisions"}],"predecessor-version":[{"id":128958,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/posts\/128951\/revisions\/128958"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/media\/128952"}],"wp:attachment":[{"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/media?parent=128951"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/categories?post=128951"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/kamucalisani.net\/index.php\/wp-json\/wp\/v2\/tags?post=128951"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}