Text Recognition in Complex Scenes Using Residual Networks and Multi-Feature Pyramid Structures

Scene Text Recognition (STR) is a crucial task in computer vision, widely used in autonomous driving, assistive reading, and smart surveillance. Despite advances in deep learning, recognizing text in low resolution, distorted, and occluded conditions remains challenging. To address this, we propose Residual Multi-Feature Pyramid Network (RMFPN), a novel feature fusion framework that integrates multi-scale visual and semantic features. RMFPN employs ResNet-50 as the backbone and utilizes a cascade and convolutional feature pyramid to enhance robustness in complex environments. Additionally, a language-aware inference module improves character dependency modeling while preserving visual integrity. Experiments on six STR benchmarks (IC13, SVT, IIIT5K, IC15, SVTP, and CUTE80) show that RMFPN achieves state-of-the-art (SOTA) performance, outperforming existing models, especially on irregular text datasets. Moreover, RMFPN adapts seamlessly to new datasets without fine-tuning. The proposed RMFPN framework provides an effective and robust solution for real-world STR, balancing feature extraction and semantic reasoning, setting a new benchmark in scene text recognition of complex.