Towards the Unseen: Iterative Text Recognition by Distilling from Errors

We introduce a meta-framework that could be incorporated on the top of most of the SOTA text recognition frameworks. Along with seminal work like ASTER by Shi et al. [16], we consider some latest text recognition frameworks like Show, Attend and Read (SAR) [12] and SCATTER [14] as our baseline, upon which we validate our iterative framework. SAR extends the work of Shi et al. using 2D attention mechanism to eliminate the rectification network. On the other side, SCATTER additionally couples multiple BLSTM encoders for richer context modelling, followed by a gating mechanism to balance between context rich information and backbone CNN features. Here, we additionally add the results using lexicons for different datasets under both for DS and CS setups. Furthermore, we add graphs to focus on by what margin our framework can provide gain over respective baseline SOTA frameworks under both DS and CS setups. Moreover, recognising rarely/unseen words being the major focus of this work, it can seen clearly seen that we achieve a rather significant performance gain on both STR (5-7%) and HTR (10-12%) tasks under DS set up. Nevertheless, we obtain a reasonable gain under CS set up unanimously over all the studied baseline SOTA frameworks.