Document Reader using Deep Learning with a Small Dataset
summary
Real-time document reader built with TensorFlow research models uses object detection to crop fields then an LSTM+Inception sequence model to read text, producing structured JSON. Original labeled set of ~700 images expands via augmentation to ~5,000 instances using Gaussian blur/noise, brightness changes, piecewise affine warps and image inpainting to remove and repaste synthetic names. Object-detection AP rises from ~60–70% to about 95% after augmentation; number recognition reaches ≈95% while names linger near ≈75%. Deployment uses Docker, Flask and AWS for a simple demo endpoint.
AI-written summary · about
chapters (6)
related talks
42mUsing Deep Learning to Understand Documents
29mdeepdoctection - An open source package for document intelligence
46mImage Recognition with Deep Learning From prototype to production
30mStarting deep learning projects without enough labelled data
01h 15mTrain Object Detection with Small Datasets
01h 00mComplex Data Ingestion with Open Source AI
41mOCR in the Wild World of Social Media
01h 07mDeep Learning with TensorFlow