A Deep Learning-based Deraining Approach using Frequency Domain Loss

• Main Authors: Yamamoto, Yusuke, Li, Yinhao, Taga, Hiroshi, Iwasa, Koki, Shichikawa, Ryuichi, Suganami, Makoto, Nakamoto, Kazuhiro, Chen, Yen-Wei
• Format: Conference Proceeding
• Language: English
• Subjects: Consumer electronics; Deep learning; Discrete cosine transforms; Frequency-domain analysis; image enhancement; Image quality; image quality restoration; Image restoration; Learning systems; Noise; Rain; Rain streak noise removal
• ISSN: 2693-0854
• DOI: 10.1109/GCCE62371.2024.10760976

Rain streak noise degrades image quality, making its removal essential for restoring image clarity. Deep learning-based methods can effectively remove rain streak noise by learning from public image datasets with many raining and clear data pairs. However, since rain streak noise affects the frequency components of an image, it is necessary to focus on these components to further enhance the performance of deep learning methods. To investigate the impact on frequency components, we analyzed images with added rain streak noise using Discrete Cosine Transform (DCT). This analysis reveals that rain streak noise affects specific frequency components. Based on these findings, we propose a new loss function called "Frequency Domain Loss." This loss function aims to effectively eliminate the frequency-based impact of rain streak noise by calculating the difference between the DCT of the network output image and the ground truth image. Additionally, to focus the loss function on specific differences, a cropping operation is introduced to extract only certain frequency components. Experimental results demonstrate that the proposed loss function significantly improves the quality of restored images.