Few-shot learning is crucial in machine learning and computer vision. It enables models to recognize new objects with limited labeled data, addressing the challenge of data scarcity and expanding the application of machine learning to domains with scarce data. Previous methods built metric space using labeled data from the base set and then classified queried images from the novel set by finding the nearest class prototype. However, due to the presence of poor-quality data in the novel set, the class prototype often exhibits instability. In response to this challenge, this paper proposes a Semantic Conditional Translation Network for reconstructing stable class prototypes. Specifically, images are first divided into edge domain (i.e., images at the cluster edge) and prototype domain (i.e., images at the cluster center). Then, an Enhanced Generative Adversarial Network is introduced to learn the translation from edge toward prototype, where a Non-parametric Classification Regularizer is designed to enlarge the discriminability of the translated samples. Meanwhile, class definitions are exploited as semantics providing precise descriptions and enhancing translation performance. Experimental results demonstrate that the proposed method obtains competitive results on four benchmark datasets.