Toward Improving Arbitrary-Shaped Text Detection With Boundary Adaptation in Noisy Scene Images

Accurate detection of arbitrary-shaped text in natural scenes remains a challenging problem, particularly under real-world noise conditions such as blur, occlusion, uneven lighting, and cluttered backgrounds. These factors degrade local features and boundary precision, hindering the performance of existing contour-based methods. To address this, we propose Ray Contour Net (RC-Net), a robust framework for boundary adaptation in noisy environments. RC-Net incorporates a Noise- Aware Attention Module to obtain noise-robust features and an Active Contour Point Initialization Module that emits contour points in a polar coordinate system anchored at local text centers. These are refined by a Deformation Module using lightweight transformers to ensure shape-aware, noise-tolerant adaptation. Additionally, we introduce a Tightness Evaluation Loss (TEL) to enforce precise boundary predictions. Extensive experiments on six benchmark datasets including Total-Text, CTW-1500, MSRATD500, ICDAR 2015, ICDAR2019 ArT, and ICDAR2019 MLT demonstrate consistent F-measure improvements of 2% to 3% over strong baselines, validating the effectiveness and robustness of our method.