Comparative Study on Transfer Learning for Object Classification and Detection

The recent development of deep neural learning achieved remarkable breakthroughs in object classification and detection. Deep learning has the capability of learning features automatically from data using general-purpose learning procedures. However, because deep neural networks require large amounts of data to train the parameters in the network, it is challenging to develop any object classification or detection system with a relatively small dataset. Transfer learning is an important machine learning technique that transfers the learned features in a pre-trained Convolution Neural Network (CNN) model into a new system. In this study, current state-of-the-art CNN models are reviewed in their architectures and characteristics. For the comparative study of transfer learning, the object classification and the detection systems are implemented using transfer learning with six state-of-the-art CNN models. The object classification model has achieved an accuracy of 97.01% for the three-class classification task using transfer learning. Furthermore, six different Faster R-CNN architectures are implemented for object detection. The performances of the different transferred models are compared in terms of the accuracy and the deploying speed of the new model. Experiments show that transfer learning saves training time and achieves accurate performance by fine-tuning the pre-existing deep learning model.