A hybrid ConvNeXt-BiLSTM framework for robust scene text recognition

Scene Text Recognition (STR) is a fundamental computer vision task with broad applications in autonomous navigation, document digitization, and assistive technologies. However, traditional STR models often rely heavily on large synthetic datasets due to the scarcity of annotated real-world data, which limits their generalization in complex environments. To address this challenge, this study proposes a ConvNeXt-based deep learning framework that integrates Convolutional Network Next (ConvNeXt) for robust feature extraction with Bidirectional Long Short-Term Memory (BiLSTM) networks for effective sequence modeling. The framework incorporates label smoothing and focal loss to enhance training stability and alleviate class imbalance and overconfidence issues. Training is conducted in two stages: pre-training on synthetic datasets (MJSynth and SynthText) followed by fine-tuning on diverse real-world datasets, including IC13, IC15, RCTW, ArT, LSVT, MLT19, ReCTS, COCO-Text, Uber-Text, TextOCR, OpenVINO, and a subset of Union14M-L. Experimental results demonstrate that the proposed model achieves an average accuracy of 94.71% over six standard STR benchmarks (IIIT5k, SVT, IC13, IC15, SVTP, and CUTE80) when trained on both synthetic and real datasets, surpassing the 89.1% accuracy achieved using synthetic data alone on the same benchmarks, and outperforming state-of-the-art methods trained under comparable data conditions. The integration of ConvNeXt, BiLSTM, advanced loss functions, and heterogeneous datasets substantially improve STR performance, particularly under challenging conditions involving irregular text layouts, multilingual content, and complex backgrounds. Furthermore, the complete recognition pipeline achieves 20.3 M parameters, 1.9 GFLOPs, and an inference latency of 2.638 ms per image, demonstrating the practical suitability of the proposed framework for real-time deployment.