Two-Stage Geometric Rectification with Attention for Scene Text Recognition

Recognizing text in natural scene images remains a challenging task due to irregular distortions such as curvature, rotation, and perspective transformations.To address these challenges, we propose a Two-Level Rectification Network (TRAN) that performs progressive correction prior to recognition.The framework integrates a Geometry-Level Rectification Network (GEO) based on Thin-Plate Spline (TPS) transformation to correct global geometric distortions, followed by a Pixel-Level Rectification Network (PIX) that refines residual local misalignment using dense offset prediction.To further enhance feature representation, we introduce a Channel-Kernel Attention Unit (CKUnit) that adaptively reweights feature channels and convolutional kernels, enabling dynamic receptive field adjustment.The rectified features are then processed through a ResNet-based feature extractor and a Bidirectional LSTM (BiLSTM) sequence modeling module, followed by Connectionist Temporal Classification (CTC) decoding for final text recognition.The entire architecture is implemented in PyTorch based on the ClovaAI deep-text-recognition-benchmark framework and initialized with pretrained TPS-ResNet-BiLSTM weights for stable convergence.Experimental evaluations on standard benchmarks including IIIT5K, SVT, ICDAR2013, ICDAR2015, SVTP, and CUTE80 demonstrate that the proposed two-stage rectification strategy achieves improved robustness and recognition accuracy compared to single-stage rectification approaches.