Currently, Vietnam has approximately 2 million blind people, which is not a small number. Therefore, developing intelligent embedded systems to implement solutions that assist the blind in their daily lives and help them better integrate into society is a topic of great interest to many researchers. One such problem is the recognition of Vietnamese currency using computer vision. Implementing this problem on mobile hardware platforms (with low configurations) requires a deep learning architecture that not only achieves high recognition accuracy but is also small in size. In this paper, we propose a solution to find a deep learning model with very high recognition accuracy (98%) on a dataset of over 9,683 images of Vietnamese currency. This model is very compact in size (512.8 KB). The results of Vietnamese currency recognition using a real camera mounted on a Raspberry Pi 4 also show an accuracy of up to 91.25%. This confirms that our solution can be fully applied to similar problems, such as recognizing objects encountered on the road, recognizing text on packaging, etc., making the system that supports the blind even more complete.