Models based on Transformers have demonstrated cutting-edge performance, albeit with a pronounced dependence on parallel corpora in the domain of neural machine translation (NMT). For low-resource languages, insufficient parallel data limits translation performance. To alleviate this issue, we present self-sampling, an efficient data augmentation (DA) method that leverages the capabilities of the model to generate pseudo sentence pairs, thereby enhancing performance for low-resource scenarios. Specifically, the first step involves constructing pseudo-monolingual sentences by employing the target-to-source NMT model to translate the target sentences from the training data-set. Furthermore, a filtering strategy is proposed to filter out sentences with high consistency compared with original sentences in the bilingual corpus. Finally, two novel DA strategies are presented, aimed at generating pseudo bilingual sentences to extend the training data-set for translation tasks in low-resource scenarios. Across a diverse range of low-resource scenarios, experimental findings reveal a substantial enhancement in translation quality, achieving a maximum BLEU score improvement of 0.64 over a robust baseline. The proposed method effectively addresses the difficulties arising from data scarcity. It yields significantly superior results compared to various alternative data augmentation methods.