zihaomu/paddleocr

PaddleOCR is a tool that reads text out of images. You give it a photo of a receipt, a screenshot, a scanned document, or a street sign, and it gives you back the actual text, editable, searchable, and ready to use in your app. It supports over 80 languages, including Chinese, English, Japanese, Korean, and Hindi, so it works for a genuinely global audience rather than being limited to one or two languages. At a high level, it works in stages: first it finds where text appears in an image, then it figures out the orientation (whether text is sideways or upside down), and finally it translates those image regions into actual characters. Beyond just extracting plain text, a companion feature called PP-Structure can understand the layout of a document, it can identify tables, pull out key information like names and dates from forms, and even convert a PDF into an editable Word document. The project also includes tools to help you label your own training data and generate synthetic text images if you want to build custom models for a specific use case. The people who would get the most out of this are builders who need text extraction in their product without paying for a cloud API. A founder building an expense-tracking app could use it to scan receipts. A PM working on document automation could use it to convert scanned contracts into searchable text. A team working on a mobile app could deploy a lightweight version (as small as 9, 16 megabytes) directly on phones, which means it works offline and doesn't send user images to a server. What's notable is the emphasis on keeping things small and fast. The "ultra-lightweight" models are designed to run on everything from full servers to mobile phones and embedded devices, which is unusual, most OCR tools assume you have a server doing the heavy lifting. The project also provides pre-trained models you can use immediately, so you don't need to train anything from scratch to get started. There's even a JavaScript version that runs in the browser, which opens up web-based use cases.