Today, AI has become one of the biggest buzzwords, but we have still not yet seen deployment at scale, at least not at the scale some of us have expected. OCR, or optical character recognition, may well be an exception.

Perhaps not as sexy or phenomenal as facial recognition or autonomous driving, or other bleeding-edge AI technologies, OCR has turned out to be a very promising AI application. OCR is useful in a very wide range of scenarios, such as logistics, manufacturing, finance, insurance, healthcare, education, government, law enforcement, and Internet.

HUAWEI CLOUD OCR features:

- High accuracy: 99 percent accuracy when reading certificates and receipts

- High adaptability: Works well even when text is misaligned, overlapping, or at an odd angle, or when text is partially obscured by official stamps, or because it is printed on a reflective surface.

- High availability: The API can be called billions of times or even more each month.

- Multi-language support: OCR recognizes multiple languages, including Thai and Burmese.

Typical Scenarios:

Recognition of Thai and Burmese ID cards. The ID cards used in many countries do not have embedded microchips, so the information on them needs to be entered manually. OCR automates this process, improving the efficiency of relevant government agencies.

Image recognition for e-commerce. Such images include web page screenshots (both PC and mobile phone), social or messaging app screenshots, ad pictures, and posters.

HUAWEI CLOUD OCR can recognize various types of images. Customers can use the results of OCR to automatically filter images by forbidden words or do some data mining. They can also extract contact information, phone numbers, and location information from images. Delivery service apps use OCR to extract address information from photos sent from users' mobile phones.

The future for OCR:

The application of OCR in an even wider range relies on further improvement of the technology, for example, the merging of different APIs for recognizing certificates and receipts, and the continuous acceleration of training and inference speeds. Faster training means faster product iteration, and faster inference means significant cost reduction for product development. Small-sample learning, unsupervised learning, and transfer learning allow customers to quickly iterate products even when they do not have any data or have only small amounts of data.