A Transfer Learning Method Based on Residual Block

In order to obtain high image representations in limited amount of datasets, a transfer learning method based on residual block is proposed, In this method, we follow a transfer learning approach by increasing the number of layers of the network to extract the higher order statistical features of the image. The main idea is to conduct feature transfer by means of ResNet (Deep Residual Network) model with setting ImageNet dataset as source domain. Firstly, all image data are preprocessed with data enhancement. Then, on the basis of modifying the source model's fully-connected classification layer, the adjustment module-residual block is added to the end of the network. Finally, after training the adjustment module, the deep model is achieved. Through transfer learning and deep feature etraction, the capability of feature recognition that impacted by content differences between source domain and target domain will be improved, Experiments show that our method achieves 97.98% accuracy on MNIST dataset and 90.45% accuracy on CIFAR-IO dataset, respectrvely. The expertmental results demonstrate that the performance of our proposed method is significantly better than the existing transfer learning methods.