How To Efficiently Increase Resolution in Neural OCR Models

Abstract

Modern CRNN OCR models require a fixed line height for all images, and it is known that, up to a point, increasing this input resolution improves recognition performance. However, doing so by simply increasing the line height of input images without changing the CRNN architecture has a large cost in memory and computation (they both scale O(n2) w.r.t. the input line height).We introduce a few very small convolutional and max pooling layers to a CRNN model to rapidly downsample high resolution images to a more manageable resolution before passing off to the "base" CRNN model. Doing this greatly improves recognition performance with a very modest increase in computation and memory requirements. We show a 33% relative improvement in WER, from 8.8% to 5.9% when increasing the input resolution from 30px line height to 240px line height on Open-HART/MADCAT Arabic handwriting data.This is a new state of the art result on Arabic handwriting, and the large improvement from an already strong baseline shows the impact of this technique.