OCR
Optical Character Recognition (OCR) is a technology that converts printed or handwritten text in images into editable, searchable digital text. Its core principle relies on image processing and pattern recognition algorithms to identify character information in scanned documents, photographs, PDFs, and other image formats. OCR is widely used in document digitization, receipt recognition, and identity document information extraction, representing an important intersection of computer vision and natural language processing.
Core Facts
GPT-4V、Claude 3、Gemini是现代多模态大模型的代表，能够同时处理文本、图像等多种数据类型
Modern OCR engines such as Tesseract and ABBYY FineReader have already achieved accuracy rates above 95% for plain text recognition.
Automated License Plate Recognition (ALPR) systems automatically capture the license plates of passing vehicles via high-resolution cameras and store information such as plate numbers, timestamps, and geographic locations in a database.
Timeline (last 90 days)
A standard chunker will accept scanned documents wholesale, splitting pages, computing embeddings for the text output by the OCR layer and storing them in the vector database, so that embeddings of noisy text still land near the correct region of the vector space, chunk cosine similarity during retrieval looks normal, and the generator gives confident answers on a contaminated premise, with the error happening silently.
Modern chunkers typically read the plain-text layer output by OCR directly and discard coordinate information, causing cells in two adjacent columns of a table to possibly be concatenated into a meaningless string in the OCR plain-text output, which the chunker then treats as a semantically complete paragraph and sends off for embedding.
Scanned PDFs have no text layer, so ordinary PDF converters cannot extract text from them and must go through an OCR path.
A pragmatic minimum viable solution includes: periodically capturing the chat area via OCR to recognize text, maintaining a FAQ keyword mapping table, judging with string matching or lightweight intent classification, simulating keyboard input to send preset answers, and setting a frequency cap.
The first step in building a knowledge base is data conversion, for example using OCR to recognize the text of images in documents and convert it into searchable text.
Not all manual data entry is worth automating; the criterion is frequency and time cost — if a conversion task occurs daily or weekly and takes more than a few minutes each time, it is worth evaluating an OCR solution.
OCR technology began in the 1950s, when early systems could only recognize specific fonts and fixed layouts.
OCR is not 100% accurate, so validation should be made a fixed part of the process rather than an optional one, in order to ensure data quality.
Low-quality images, skewed captures, handwriting and stamp occlusion are common sources of OCR error, and preprocessing steps such as denoising, binarization, perspective correction and rotation alignment can often significantly improve results.
The coordinate information in OCR output (bounding boxes) is key to reconstructing table structure, and by analyzing the positional relationships of text boxes one can infer row and column membership and generate structured JSON or CSV.
40 more timeline events
All Facts (20)
GPT-4V、Claude 3、Gemini是现代多模态大模型的代表，能够同时处理文本、图像等多种数据类型
90%VerifiedModern OCR engines such as Tesseract and ABBYY FineReader have already achieved accuracy rates above 95% for plain text recognition.
80%VerifiedAutomated License Plate Recognition (ALPR) systems automatically capture the license plates of passing vehicles via high-resolution cameras and store information such as plate numbers, timestamps, and geographic locations in a database.
65%VerifiedPDF is fundamentally a print-oriented format, focused on where pixels appear on a page rather than the logical structure of text.
65%VerifiedTraditional OCR technology excels at converting image text into machine-readable text, but has limited ability to understand the semantic structure of documents.
65%VerifiedOCR（光学字符识别）是一项将图像中的印刷或手写文字转化为可编辑数字文本的技术，起源于20世纪50年代
65%VerifiedOCR solutions based on multimodal large models often perform better than traditional OCR solutions when handling complex layouts, handwriting, and multilingual mixed typesetting scenarios.
65%VerifiedCurrent mainstream OCR solutions have achieved recognition accuracy exceeding 99% on standard printed text.
65%UnverifiedALPR technology combines Optical Character Recognition (OCR) with infrared imaging to read license plate information under various lighting and weather conditions
90%Unverified多模态大模型的推理成本远高于传统OCR引擎
90%Unverified该工具集成了OCR光学字符识别功能，能够自动识别屏幕上的文字内容
85%UnverifiedPrescription parsing typically follows a two-stage process: an OCR engine converts the prescription image into raw text, then a large language model performs semantic extraction
80%UnverifiedAI prescription parsing typically combines OCR with LLM-based information extraction to convert prescription details into structured medication reminder plans
80%Unverified该项目使用OCR和图像识别替代传统的内存读取方式进行游戏状态感知，优势在于不需要逆向分析游戏内存结构，也不容易触发反作弊系统的检测
80%UnverifiedDetecting hallucinations requires a dedicated evaluation mechanism independent of accuracy, such as comparing model outputs with the literal correspondence to the original OCR text.
70%UnverifiedOCR accuracy drops significantly for artistic fonts, severely distorted text, or designs where text blends into the background.
60%UnverifiedMultimodal large models can infer partially occluded text based on contextual semantics, with recognition accuracy and robustness far exceeding traditional OCR solutions.
60%UnverifiedOCR technology can be traced back to photoelectric voting machines and reading devices for the blind in the 1920s.
60%UnverifiedText-image semantic alignment established by technologies such as CLIP enables models to map abstract style descriptions to concrete visual feature spaces
60%UnverifiedModern OCR systems typically consist of three core modules: text detection, text direction classification, and text recognition.
60%