LCSTR: Scene Text Recognition with Large Convolutional Kernels

Abstract

The task of scene text recognition involves processing information from two modalities: images and text, thereby requiring models to have the ability to extract features from images and model sequences simultaneously. Although linguistic knowledge greatly aids scene text recognition tasks, the extensive use of language models in sequence modeling and model prediction stages in recent years has made model architectures increasingly complex and inefficient. In this paper, we propose LCSTR, a pure convolutional visual model that can complete text recognition without the need for attention mechanisms or language models. This approach applies large kernels to text recognition tasks for the first time, extracting word-level text information through large text-aware blocks, capturing long-range dependencies between characters, and using small text-aware blocks to obtain local features within characters. Experiments show that this model strikes a good trade-off between accuracy and speed, achieving notable results on seven public benchmarks, validating the generalizability and effectiveness of this method. Furthermore, owing to the absence of a language module, this model demonstrates remarkable accuracy even in limited sample scenarios, and the lightweight and low computational overhead features make it suitable for engineering applications.