One-shot object recognition is a challenging task for deep neural networks. It becomes more difficult when there is a domain shift between support and query samples. The generalization of a deep model on an unknown domain with different distributions is another problematic task in on-shot recognition. This work introduces a novel deep network architecture named SENet and a multi-step meta-training strategy to overcome the aforementioned problems. To the best of our knowledge, this work outperforms all state-of-the-art models in one-shot traffic sign recognition and one-shot logo identification by superior results. Although the SENet model achieves comparable results for the traditional GTSRB benchmark, it has about 6X fewer parameters than its competitors. The smaller size of the SENet architecture enables it to be used in resource-constrained devices in many applications such as smart vehicles. The experimental results depict that our proposed SENet performance is ameliorated by large margins, with up to 20% accuracy for one-shot classification and 30% area under the curve (AUC) for image retrieval tasks.