.OCR a scenario
1. four categories:
digital primary categories:
Taobao FIG digital goods native class most representative character of FIG.
Features: 1) the most complex and diverse: a variety of fonts, background, permutations, combinations, etc. (MTWI Challenge - the biggest contest OCR).
2) The most valuable: the product information carrier
3) The maximum amount of pictures: one hundred billion images, constantly updated daily.
Document type:
Document Class OCR demand is very wide, involving a variety of public service scenarios.
Features: 1) 100% recognition rate: the human input accuracy was 98%, and explore the limits of AI knowledge; 2) ease of use: fully functional, close to the business needs; 3) commercial applications: document class commercial maturity.
Form class pictures: pictures form class OCR value is very large, very challenging.
Features: 1) Scene & Data: Data has privacy, a typical application scenarios precipitation technical capacity; 2) Product versatility: expert knowledge + template = text comprehension, a set of solutions to hundreds of types. 3) Business Value: industry and scene depth access, AI capabilities to improve industry data flow. (Table recognition camera offers customized and structured cloud services)
natural scene categories: key directions of academic research OCR.
Features: 1) Data: no specific data type definitions, such as street shooting data; 2) Technical difficulties: uncertainty, complicated interference environment is to locate and identify the nature difficulty; 3) the commercial value: the market has great potential, such as: license plate recognition, video surveillance, automatic pilot. (Leading technical capabilities, landing in the industry)
2.OCR algorithm:
Algorithm capacity:
two kinds of core algorithm capability: 1) General character recognition; 2) of the general structure (character recognition is an infrastructure of)
basic algorithms: positioning text, character recognition
1. Text Location: positioning the character position in the image characterization rows.
Characterized by questions: Background characteristics such interference problems, the depth of learning problems can be better solved features.
Scale problems: positioning objects common problems, character height range 8-300 pixels
in rows problems: positioning text specific problem.
(1) Scale issues: common problems positioning objects
(2) in rows problems: positioning issues specific character
2. character recognition
on the basis of the text positioned on text recognition, and also outputs the position and the word recognition rate for understanding the text.