While modern communication technologies have made online information more accessible to Cambodians, the lack of robust Khmer machine translation (MT) tools continues to limit the serviceability of content that is primarily available in English. Furthermore, the development of high-quality Khmer-English MT systems is hindered by the scarcity of parallel corpora, as Khmer remains a low-resource language. This paper proposes a pragmatic, data-driven approach to enhancing Khmer-English MT by leveraging document analysis techniques to automatically mine high-quality parallel corpora. Our method combines document layout analysis, text recognition, and multilingual text embedding alignment to extract parallel Khmer-English sentence pairs from diverse sources, including news articles, press releases, and books. Using this approach, we compile a dataset of 32,180 high-quality Khmer-English sentence pairs, which will be publicly released. Finetuning an MT model with this additional data yields state-of-the-art (SoTA) translation performance across semantic and non-semantic evaluation metrics. Specifically, the finetuned model improves translation quality by 62% and 76% for the Khmer \(\rightarrow \) English and English \(\rightarrow \) Khmer directions, respectively, in comparison with the baseline. This highlights the importance of the document analysis techniques in enhancing low-resource language machine translation.