OCR Academic Frontiers and Industrial Application Summit Forum
Related agenda: https://mp.weixin.qq.com/s/LYoKHFad9D-gjhGlVF3Czg
Advertising OCR Technology Research and Application-Tencent
Video making ASR, ocr get subtitles
computer animation CG
OCR Practice and Technology Innovation - Ant
-
loss optimization
-
data synthesis
Contrasting the learning method, what is a positive sample and what is a negative sample.
Generative self-supervised learning.
Connected character blocks have a similar style
Decoding content and style
Additional application scenarios: cross-language editing, font transformation
Heat map to view the effect
- Multimodal document image understanding
Representation of positional relationship, up-down, left-right, high-dimensional relationship
pre-fusion and post-fusion
- prior knowledge representation
New Thoughts on Handwritten Character Recognition——Data, Method and Application-Jin Lianwen/South China University of Technology
A few thousand-2w level key
The latter two categories are the mainstream
Based on the Gan learning style library, the learning style
The naturalness of continuous strokes
autoencoder
GLRNet, one-dimensional convolution is a local feature, and the encoder is a global feature
Semantic module CTC
Generation and Detection of Tampered Text Images-University of Science and Technology of China
The texture recognition effect of complex background is not good
Frame manipulation
Tampering of Curved Text
Open Set Text Recognition: Concept, Framework, Algorithm and Application-Beijing University of Science and Technology
The traditional character is not necessarily itself, but has both front and back features
In the case of open questions, you should focus on your own identification features to avoid being corrected by mistakes
New Advances in Text-Oriented Graphics and Image Generation Technology-Peking University
Full-stack R&D of OCR and Practice in Industry Scenarios – Huawei
Tampering Detection of Qualification Documents and Certificate Images and Application in Digital Economy Scenarios-Alibaba
Dozens of types of documents
- Two categories
- true and false identification
OCR Industrialization Application Practice-Shanghai Hehe Information
Image preprocessing, layout analysis and restoration are more important
all in one model
Stamp and text are layered
Research progress on end-to-end mathematical formula recognition combined with domain knowledge-University of Science and Technology of China
Research on low-quality scene text recognition technology-Institute of Information Technology, Chinese Academy of Sciences
From PaddleOCR to see the innovative direction of OCR industry landing-Baidu
Video OCR Technology and Application - Byte
The text track is query
sliding window
Detector missed detection, false detection, mismatch, etc.
track shaper
Research Progress of OCR in Vertical Application—Hikvision
color correction
Transformer does not work well on some hardware platforms, and the large model is not very friendly in practical applications.
Document classes should use multimodal techniques