TEACH: Text Encoding as Curriculum Hints for Scene Text Recognition

TL;DR

TEACH introduces a curriculum learning approach for scene text recognition by gradually reducing reliance on ground-truth text during training, leading to improved accuracy without extra inference costs.

Contribution

It presents a novel training paradigm that encodes target labels as auxiliary inputs and employs loss-aware masking, enhancing model robustness without external pretraining.

Findings

Consistent accuracy improvements across benchmarks.

Enhanced robustness in challenging conditions.

No additional inference overhead.

Abstract

Scene Text Recognition (STR) remains a challenging task due to complex visual appearances and limited semantic priors. We propose TEACH, a novel training paradigm that injects ground-truth text into the model as auxiliary input and progressively reduces its influence during training. By encoding target labels into the embedding space and applying loss-aware masking, TEACH simulates a curriculum learning process that guides the model from label-dependent learning to fully visual recognition. Unlike language model-based approaches, TEACH requires no external pretraining and introduces no inference overhead. It is model-agnostic and can be seamlessly integrated into existing encoder-decoder frameworks. Extensive experiments across multiple public benchmarks show that models trained with TEACH achieve consistently improved accuracy, especially under challenging conditions, validating its robustness and general applicability.