Towards Robust Scene Text Recognition: A Dual Correction Mechanism with Deformable Alignment

Abstract

Scene Text Recognition (STR) faces significant challenges under complex degradation conditions, such as distortion, occlusion, and semantic ambiguity. Most existing methods rely heavily on language priors for correction, but effectively constructing language rules remains a complex problem. This paper addresses two key challenges: (1) The over-correction behavior of language models, particularly on semantically deficient input, can result in both recognition errors and loss of critical information. (2) Character misalignment in visual features, which affects recognition accuracy. To address these problems, we propose a Deformable-Alignment-based Dual Correction Mechanism (DADCM) for STR. Our method includes the following key components: (1) We propose a visually guided and language-assisted correction strategy. A dynamic confidence threshold is used to control the degree of language model intervention. (2) We designed a visual backbone network called SCRTNet. The net enhances key text regions through a channel attention module (SENet) and applies deformable convolution (DCNv4) in deep layers to better model distorted or curved text. (3) We propose a deformable alignment module (DAM). The module combines Gumbel-Softmax-based anchor sampling and geometry-aware self-attention to improve character alignment. Experiments on multiple benchmark datasets demonstrate the superiority of our approach. Especially on the Union14M-Benchmark, where the recognition accuracy surpasses previous methods by 1.1%, 1.6%, 3.0%, and 1.3% on the Curved, Multi-Oriented, Contextless, and General subsets, respectively.