Residual Training for Super Resolution of Solar Dynamics Observatory and Helioseismic Magnetic Imager Magnetogram

다국어 초록 (Multilingual Abstract)

Super-resolution (SR) is a deep learning technique of enhancing the resolution of images or videos, and it is widely used in various fields such as video restoration, improving CCTV footage, and medical imaging. While performance has steadily improved using techniques such as generative adversarial networks (GANs) and diffusion models, several limitations remain. Super-resolution models, which are based on supervised learning, require a large amount of data and the performance of the model declines significantly when applied to data domains that differ from those used during training. To address this issue, transfer learning is often utilized in convolution neural network (CNN) models. However, this technique is predominantly applied in classification models and has not been utilized in models aiming at improving resolution. Then to solve this problem, we proposed a novel transfer learning method named Residual Training, which is applicable to super-resolution models. This method operates by utilizing the output of a pre-trained model as input for a residual training model, thereby improving the detail of images. This approach enables the enhancement of image quality while maintaining the original model’s structure and parameters. For validation of the proposed method, we used the Helioseismic and Magnetic Imager (HMI) images, specifically the Magnetogram images of the Sun, as our dataset.

목차 (Table of Contents)

• Abstract II
• 1. Introduction 1
• 2. Methods 3
• 2.1 Data SDO/HMI Magnetogram 3
• 2.2 Pre-trained Super Resolution Models 3
• 2.3 Proposed method 5
• 3. Assessment 10
• 3.1 Image Quality Assessment, IQA 10
• 3.2 Ten-fold cross validation 14
• 4. Results 15
• 4.1 Analysis of SR image and its residual 15
• 4.2 Assessment 16
• 5. Discussion 20
• 6. Conclusion 21