An efficient dRED-TL-GAN model for image denoising†
Ji Hye Heo1 · Dong Hoon Lim2
12Department of Information Statistics, Gyeongsang National University
Correspondence to: † This work was supported by the National Research Foundation of Korea (NRF) grant funded by the Korea government (MSIT) (No. NRF-2022R1A2C1011140).
1 Master’s graduate, Department of Information Statistics, Gyeongsang National University, Jinju 52828, Korea.
2 Professor, Department of Information Statistics and RINS, Gyeongsang National University, Jinju 52828, Korea. E-mail: dhlim@gnu.ac.kr
1 Master’s graduate, Department of Information Statistics, Gyeongsang National University, Jinju 52828, Korea.
2 Professor, Department of Information Statistics and RINS, Gyeongsang National University, Jinju 52828, Korea. E-mail: dhlim@gnu.ac.kr
Received April 3, 2024; Revised May 20, 2024; Accepted May 22, 2024.
This is an Open Access article distributed under the terms of the Creative Commons Attribution Non-Commercial License (http://creativecommons.org/licenses/by-nc/3.0) which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.
- Abstract
- Noise in images not only causes visual distortion or inconvenience, but also reduces performance in the imaging system, so image denoising is an important preprocessing process in image processing. In this paper, we propose a dRED-TL-GAN model based on DCGAN, derived from GAN, to remove noise from images. The generator of the dRED-TL-GAN model is a deformable RED structure consisting of an encoder-decoder structure, and the discriminator is a transfer learning-based structure. Here, the deformable RED structure used deformable convolution in the encoder’s convolution layer to remove noise by considering the characteristics of the image, and used the ResNet-18 model in the discriminator to increase learning speed and classification accuracy. To evaluate the performance of the proposed dRED-TL-GAN model, traditional filters including Mean filter, Median filter, and BM3D filter, and existing deep learning models including DnCNN model, RED-CNN model, and DCGAN model were considered. An performance experiment was conducted on face images damaged by various noises, namely Gaussian noise, Poisson noise, and Speckle noise. The performance experiment consists of qualitative and quantitative evaluations. First, in the qualitative evaluation, spatial filters including the Mean filter, Median filter, and BM3D filter generally remained noisy and resulted in blurry results, and the proposed dRED -TL-GAN model obtained clearer images with edges than other deep learning models. Additionally, in a quantitative evaluation using PSNR, MSE, and SSIM metrics, the dRED-TL-GAN model performs well under all noises considered and on all evaluation metrics.
- Keywords : Deep learning, deformable convolution, dRED-TL-GAN, GAN, image denoising
Full Text(PDF) Free
