I don’t expect how well of the recognition accuracy after retraining, just present here how to do the retrain.

But i think you can try to do some preprocessing on image before the recognition operation, such like: noise filter, and Binarization, that may help. Otherwise you need more similar dataset to retrain it.