OCR in the Wild World of Social Media
summary
An OCR pipeline for noisy social-media images that detects text regions, crops them, and runs recognition using SSD-style detectors with a VGG16 encoder and newer segmentation detectors such as CRAFT. The team bootstraps training with synthetic data, shows that off-the-shelf English models get ~20% word accuracy on social posts while locally tuned models reach ~44–50%, and runs a production monitor processing Facebook/Instagram/Twitter streams at about 86/20/13 FPS on GPU depending on model complexity.
AI-written summary · about
chapters (8)
- 0:00OCR for social media
- 3:30OCR tech evolution and limits
- 8:53Modern text detection
- 16:21Text recognition from images
- 23:28Universal OCR pipeline
- 27:26Synthetic data generation
- 31:58Multilingual model tuning
- 36:36OCR results and future
related talks
28mScalable High Precision OCRs for Game Event Detections
22mOCR, Recognition and Information Extraction from Images
20mScalable OCR pipelines using Python,Tensorflow & Tesseract
19mDocument Reader using Deep Learning with a Small Dataset
25mBrand recognition in real-life photos
37mCan you Read This? How I Improved Text Readability Online for Visually Impaired
53mNodeJS, ML, K8s and Unethical Face Recognition
19mRushin Shah, Engineering Manager, Facebook