Pre-training a deep neural network on the ImageNet dataset is a common\npractice for training deep learning models, and generally yields improved\nperformance and faster training times. The technique of pre-training on one\ntask and then retraining on a new one is called transfer learning. In this\npaper we analyse the effectiveness of using deep transfer learning for\ncharacter recognition tasks. We perform three sets of experiments with varying\nlevels of similarity between source and target tasks to investigate the\nbehaviour of different types of knowledge transfer. We transfer both parameters\nand features and analyse their behaviour. Our results demonstrate that no\nsignificant advantage is gained by using a transfer learning approach over a\ntraditional machine learning approach for our character recognition tasks. This\nsuggests that using transfer learning does not necessarily presuppose a better\nperforming model in all cases.\n