Scene text detection and recognition in low-resource languages such as Khmer pose significant challenges due to the scarcity of annotated datasets and the script’s inherent complexity, including stacked consonants, intricate ligatures, and context-dependent vowel diacritics. Unlike Latin-based scripts, Khmer lacks clear word boundaries and exhibits a wide variety of character combinations, making it difficult for standard OCR systems to perform reliably. While recent advances in deep learning have achieved strong performance for high-resource languages, their direct applicability to Khmer remains limited and underexplored. This study investigates the potential of cross-lingual transfer learning as a solution to bridge this gap. By leveraging pretrained models from high-resource languages such as English, we explore how existing state-of-the-art detection and recognition architectures can be adapted to the Khmer script. Specifically, we evaluate the effectiveness of fine-tuning these models, originally trained on large-scale Latin or multilingual datasets, using a limited amount of annotated Khmer data to boost accuracy and generalization.