NRSTRNet: A Novel Network for Noise-Robust Scene Text Recognition

Abstract Scene text recognition (STR) has been widely applied in industrial and commercial fields. However, existing methods still face challenges when processing images with defects such as low contrast, blur, resolution, insufficient illumination. These are common actual situations because of diverse backgrounds natural scenes limitations shooting conditions. To address these challenges, we propose a novel network for noise-robust scene (NRSTRNet), which comprehensively suppresses the noise three critical steps STR. Specifically, feature extraction stage, NRSTRNet enhances text-related features through channel spatial dimensions disregards some disturbances from non-text area, reducing redundancy input image. In context encoding fine-grained coding is proposed to effectively reduce influence previous noisy temporal on current while simultaneously impact partial overall by sharing contextual parameters. decoding self-attention module added enhance connections between different features, thereby leveraging global information obtain noise-resistant features. Through approaches, can local semantic considering information. Experimental results show that improve ability characterize images, stability under noise, achieve superior accuracy recognition. As result, our model outperforms SOTA STR models irregular benchmarks 2% average, it exceptionally robust images.