Скачать или смотреть DeepSeek-OCR: Contexts Optical Compression

This document introduces DeepSeek-OCR, a novel Vision-Language Model (VLM) designed for efficient vision-text compression using optical 2D mapping. It consists of DeepEncoder for compressing high-resolution images into vision tokens and DeepSeek3B-MoE as the decoder. DeepEncoder maintains low activations and achieves high compression ratios. Experiments demonstrate that DeepSeek-OCR achieves high OCR precision (up to 97%) when the text token count is within 10 times that of vision tokens. Even at a compression ratio of 20x, accuracy remains around 60%. The model surpasses existing OCR solutions on the OmniDocBench benchmark, utilizing fewer vision tokens. The core contribution of DeepSeek-OCR lies in its vision-text token compression capabilities and its novel DeepEncoder architecture. The model shows promise for long-context compression research and has practical applications, generating large-scale training data. Codes and models are publicly accessible.
#OCR #DeepSeek #VisionLanguageModel #Compression #DeepLearning #AI #OpticalCompression

paper - https://github.com/deepseek-ai/DeepSe...
subscribe - https://t.me/arxivpaper
donations:
USDT: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7
BTC: bc1q8972egrt38f5ye5klv3yye0996k2jjsz2zthpr
ETH: 0xAA7B976c6A9A7ccC97A3B55B7fb353b6Cc8D1ef7
SOL: DXnz1nd6oVm7evDJk25Z2wFSstEH8mcA1dzWDCVjUj9e
created with NotebookLM