Transfer Learning Techniques for Image Recognition: A Systematic Literature Review

Abstract

Deep learning is a rapidly expanding research area focusing on the use of more extended (deep) and varied neural network architectures to solve more complicated problems than traditional multi-layer perceptrons. Transfer learning is a more recent off-shoot of deep learning which focuses on using information from one machine learning task in another related task. It has primarily seen applications in image classification, for instance, when information used to recognize/classify a bicycle can be used to classify a motorcycle. In a rapidly evolving research space, it is important to summarize the research applications of different deep learning offshoots. In this regard, this paper presents the first systematic literature review particularly targeting applications of transfer learning to image recognition. We follow the standard methodology and categorize papers on the basis of more critical KPIs. Our core finding is that this particular domain is a hot area of research these days, and most applications are related to pre-trained models learnt from convolution neural network and applied to another convolution network. Also, transfer learning has lead to significant improvements in accuracy and efficiency and facilitation, as compared to learning deep models or other machine learning approaches from scratch. From our results, we propose several future directions of research.