Accurate scene text detection is crucial for most applications, including, but not limited to, automatic document processing, augmented reality, and navigation systems. However, diverse sizes, orientations, and complex backgrounds often impede detection accuracy. This paper proposes a novel framework by leveraging scale-aware data augmentation and shape similarity constraints to enhance scene text detection. The augmentation of scale awareness dynamically adjusts the text instances at multiple scales during training, improving performance in detecting small and large text. Meanwhile, the constraint on shape similarity keeps the consistency of the text shape representation for better localization and reduction of false positives. Extensive experiments on standard scene text datasets demonstrate that our approach overwhelmingly outperforms other existing methods in detection precision and recall, with the merits more evident in detecting irregularly shaped text. The framework’s robustness is spread across various real-world scenes and hence is a very promising solution for unconstrained text detection scenarios.