SVTRv2X: Enhanced scene text recognition via self-distilled mixture-of-experts

Abstract

Scene Text Recognition (STR) is a fundamental component of intelligent perception systems and plays a crucial role in a wide range of real-world applications such as autonomous driving, document understanding, and human–computer interaction. STR still faces several challenges in practical applications, including high sensitivity to spatial perturbations, limited representational capacity of lightweight Connectionist Temporal Classification(CTC)-based models, and the difficulty of handling diverse text styles within a single unified architecture. Although SVTRv2 enhances the recognition ability of CTC models through a combination of local and global mixing mechanisms, its robustness and generalization capability remain insufficient when dealing with geometric distortions, complex backgrounds, or text with large stylistic variations. To address these issues, we propose SVTRv2X, an enhanced STR framework built upon SVTRv2 that integrates three complementary improvement modules. The Jumble Module strategically rearranges input patches before the patch embedding stage, fundamentally reducing the model’s reliance on fixed spatial structures and significantly improving robustness to rotated, misaligned, and irregular text. The Self-Distillation Module transfers deep-layer knowledge to shallow features, effectively strengthening early-stage representations while maintaining lightweight inference. The Mixture-of-Experts (MoE) Module expands model capacity through sparsely activated expert networks, allowing specialized processing of different text styles without introducing substantial computational overhead. Extensive experiments demonstrate that SVTRv2X achieves state-of-the-art performance on multiple STR benchmarks, substantially advancing the model’s recognition capability in real-world scene text scenarios.