Memory Matters: Convolutional Recurrent Neural Network for Scene Text Recognition

TL;DR

This paper introduces a convolutional recurrent neural network that leverages temporal memory to improve scene text recognition without character segmentation, demonstrating superior performance over traditional methods.

Contribution

It investigates how temporal memory in CNN-RNN models enhances segmentation-free scene text recognition, utilizing LSTM units for long-term memory retention.

Findings

Outperforms traditional HMM-based methods on SVHN dataset

Effectively models appearance and temporal dynamics of scene text

Demonstrates the advantage of LSTM over short-term models

Abstract

Text recognition in natural scene is a challenging problem due to the many factors affecting text appearance. In this paper, we presents a method that directly transcribes scene text images to text without needing of sophisticated character segmentation. We leverage recent advances of deep neural networks to model the appearance of scene text images with temporal dynamics. Specifically, we integrates convolutional neural network (CNN) and recurrent neural network (RNN) which is motivated by observing the complementary modeling capabilities of the two models. The main contribution of this work is investigating how temporal memory helps in an segmentation free fashion for this specific problem. By using long short-term memory (LSTM) blocks as hidden units, our model can retain long-term memory compared with HMMs which only maintain short-term state dependences. We conduct experiments on Street View House Number dataset containing highly variable number images. The results demonstrate the superiority of the proposed method over traditional HMM based methods.