DUAL-STAGE RECTIFICATION AND ATTENTION FRAMEWORK FOR ROBUST SCENE TEXT RECOGNITION

Abstract

Recognizing scene text under irregular distortions demands robust rectification prior to decoding. We propose aTwo-Level Rectification Attention Network (TRAN) that unites a Geometry-Level Rectification Network(GEO)—leveraging thin-plate spline (TPS) warping to correct global skew and curvature—with a Pixel-LevelRectification Network (PIX) that applies fine-grained per-pixel offsets to refine local deformations. To handlediverse character scales and appearances, we introduce a Channel-Kernel Attention Unit that dynamicallyweighs feature channels and convolutional kernels. Implemented atop the ClovaAI deep-text-recognitionbenchmark framework with PyTorch and pretrained CNN–RNN backbones, TRAN demonstrates superiorrectification and recognition performance. Large-scale experiments on benchmarks with curved, rotated, andperspective-warped text demonstrate that TRAN's two-stage rectification strategy is far superior to single-stagerectification algorithms. Our results point to the potential of combining multi-level rectification with adaptiveattention as a promising direction for more robust scene text recognition in real-world applications likenavigation systems and reading aid devices.