LIU,YUAN CHEN
National Taipei University of Education Department of Computer Science Professor
Email：liu@tea.ntue.edu.tw
HSIEH,TZU HSUAN
National Taipei University of Education Department of Computer Science Student
Email：s110616041@stu.ntue.edu.tw
Abstract
With the continuous advancement of deep learning technology, many convolutional neural network architectures have been applied in superresolution tasks, but many convo-lutional neural network-based models have many parameters and depth of the model make it more and more difficult to train the model, and it is more difficult for each parameter to converge to the optimal value. Most of the research focuses on the design of the model, while ignoring the influence of hyperparameters on the model, resulting in the model not being able to exert the complete ability of the model and spending more time training. In this paper, through numerous experiments, we find out the suitable hyperpa-rameter combination for the super-resolution imaging model, optimize the training of the model without greatly changing the model architecture, and make the model have a high accuracy. Experiments have confirmed that under different magnification factors and dif-ferent model architectures, our optimized models and training methods have obtained high prediction accuracy of super-resolution imaging in various validation sets. Finally, we found a set of hyperparameters suitable for superresolution imaging, us-ing SmoothL1 loss, PReLU excitation function, RMSProp optimizer, and using post-upsampled sub-pixel convolution such a combination can make the same model architec-ture have more high accuracy.
Keywords :Artificial neural networks, Super resolution, hyperparameters