Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7, 1345-1359 2025 Publisher: Learning Gate DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate © 2025 by the author; licensee Learning Gate History: Received: 6 May 2025; Revised: 7 June 2025; Accepted: 11 June 2025; Published: 17 July 2025 * Correspondence: gulden.eleyan@acm.edu.kw Image denoising using deep learning: Comparative study Gülden Eleyan1* 1Department of Engineering and Technology, American College of the Middle East, Egaila 54200, Kuwait; gulden.eleyan@acm.edu.kw (G.E.). Abstract: Deep learning offers a flexible and effective approach to automated image denoising. This study investigated the residual learning capabilities of two recent deep learning networks: the Restoration Transformer (Restormer) network and the Deep CNN (DnCNN) network. We compared their denoising performance on the BSD68 dataset under varying levels of Gaussian noise against established algorithms such as block-matching and 3D filtering (BM3D) and trainable nonlinear reaction diffusion (TNRD). Our findings demonstrate that the Restormer algorithm excels in noise removal. This highlights the potential of transformer-based architectures in image restoration tasks, surpassing traditional methods in achieving superior denoising quality. Further research can explore the application of Restormer to other noise types and datasets. Keywords: Convolutional neural networks, Deep learning, Image denoising, Transformers. 1. Introduction Deep learning is a powerful tool that enables the automatic removal of noise from images in a flexible and effective manner. One of the most significant advantages of using deep learning for image denoising is its ability to learn and capture complex patterns in the data, unlike traditional image denoising techniques, which may not effectively remove noise in all cases. Deep learning-based approaches have been successfully applied in various fields, including low-light image enhancement, noise reduction in medical imaging, and image restoration. Noise reduction in images is the process of removing noise to improve the visual quality of an image. Noise can be introduced into an image through various means, such as capturing in low-light conditions, image compression, or transmission over a noisy channel. Removing noise from an image helps restore its original clarity and enhances its overall appearance [1]. Several techniques can be used to remove noise from images, including adaptive filters [2-5] and wavelet-based denoising [6-8]. Smoothing filters work by averaging the pixel values in an image, while median filters replace a pixel’s value with the median value of its surrounding pixels. Wavelet-based denoising employs wavelet transforms to separate an image into different frequency bands, allowing noise to be distinguished from the signal and removed. Completely eliminating noise often requires sacrificing some level of detail, but with careful tuning, it is possible to minimize detail loss while achieving effective noise reduction. Deep learning has recently emerged as a powerful tool for image noise reduction. In deep learning, a neural network is trained to learn the underlying structure of an image and predict a noise-free version from a noisy input [1]. This is typically achieved using a large image dataset containing both noisy images and their corresponding clean versions, which is used to train the network to learn the relationship between noisy and noise-free images. One of the key advantages of using deep learning for image denoising is its ability to learn and capture complex structures within data. Traditional image denoising techniques rely on fixed filtering, which may not effectively remove noise in all situations. In contrast, deep learning-based approaches can 1346 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate adapt to different types of noise and denoise images more flexibly and effectively. Additionally, deep learning models can be trained in an end-to-end manner, allowing them to learn from large datasets and automatically determine the most suitable noise reduction technique for a given task. As a result, deep learning has the potential to enhance the performance of image denoising techniques and has already been successfully applied in various scenarios. The rest of this work is structured as follows: The history and related work on image denoising are summarized in Section 2. Section 3 discusses and presents two recently introduced image denoising algorithms explored in this project—the Restoration Transformer (Restormer) network and Deep Convolutional Neural Networks (DnCNN) Residual Learning. The results and comparisons are presented and analyzed in Section 4. Finally, Section 5 concludes the project and discusses future work. 2. History and Related Work 2.1. Convolutional Neural Networks (CNNs) CNNs have achieved remarkable success in image processing due to their plug-and-play network architectures [9-14]. As a pioneer in CNN technology, LeNet [15] utilized convolutional kernels of different sizes to extract features and achieve effective image classification. However, due to the use of the Sigmoid activation function, LeNet had a slow convergence rate, which posed a limitation in real- world applications. Following LeNet, AlexNet Krizhevsky, et al. [16] and Daalah, et al. [17] became a milestone in deep learning (see Figure 1). Its success stemmed from its ability to address the overfitting problem and improve the speed of stochastic gradient descent (SGD) instead of using the Sigmoid activation function [18]. While AlexNet achieved high performance, its large convolutional kernels required significant memory usage, limiting its application in real-world scenarios such as smart cameras. Figure 1. LeNet (Top) and AlexNet (Bottom). 1347 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate To enhance performance and reduce computational costs, deeper network architectures with smaller filters were preferred. Specifically, VGG [19] stacked more convolutional layers with small kernel sizes, as illustrated in Figure 2. Figure 2. The Architecture of the VGG. As accuracy improved, research shifted toward increasing the width of networks. GoogleNet Szegedy, et al. [20] expanded the width of CNN architectures to enhance the performance of image- processing applications. Additionally, it reduced the number of parameters and computational costs by replacing large convolutional kernels with two smaller ones. The fundamental convolutional block in GoogleNet is called the Inception block (see Figure 3), and the GoogleNet architecture is shown in Figure 4. Figure 3. The Inception Block. 1348 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate Figure 4. GoogleNet Architecture. Although VGG and GoogleNet methods are effective for image applications, they come with two main disadvantages: if the network is too deep, it may suffer from vanishing or exploding gradients; if the network is too wide, it may be prone to overfitting. To overcome these issues, ResNet He, et al. [21] was proposed in 2016. To improve image recognition performance, residual learning was introduced into each block of the ResNet architecture. Figure 5 shows the structure of ResNet-18, while the basic building block of ResNet is illustrated in Figure 6. Figure 5. ResNet-18 Architecture. Figure 6. ResNet Block. Traditional machine learning algorithms have long been employed in diverse image processing applications [22-27]. For nearly a decade, deep networks have seen widespread use in real-world image 1349 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate applications, including face recognition [28-32] and medical diagnosis [33-37]. However, in many applications, such as real noisy image scenarios, the captured images are often insufficient in quality, and deep CNNs tend to show limited performance. To address this, Generative Adversarial Networks (GANs) Radford, et al. [38] were developed. GANs consist of two networks: A Generator and a Discriminator. The generator is used to produce samples based on input data, while the discriminator evaluates the authenticity of both the input samples and the generated ones. These two networks work in opposition: if the discriminator can accurately distinguish real samples and the generator can produce convincing fake ones, the model is considered successfully trained. The architecture of a GAN is illustrated in Figure 7. Due to their ability to generate complementary training examples, GANs are highly effective in working with small datasets, especially in tasks like face recognition [39] and denoising complex noisy images [40]. Figure 7. GAN Architecture. 2.2. Deep Learning for Image Denoising With the introduction of the networks mentioned above, deep learning techniques have attracted considerable attention in the field of image denoising. Researchers have explored deep neural networks to address the problem of noise removal. However, there are significant differences among various deep learning approaches to image denoising. Specifically, discriminative learning based on deep learning has proven effective in tackling the problem of Gaussian noise. In Jain and Seung [41] proposed using Convolutional Neural Networks (CNNs) for image denoising, arguing that CNNs can provide representations like, or even better than, those of Markov Random Field (MRF) models [42]. In Burger, et al. [43]. Multi-Layer Perceptrons (MLPs) were successfully applied to image denoising. In Xie, et al. [44] stacked sparse denoising autoencoders were adopted for Gaussian noise removal and achieved results comparable to K-SVD [45]. In Chen and Pock [46] a Trainable Nonlinear Reaction Diffusion (TNRD) model was proposed, which can be expressed as a feedforward deep network by unrolling a fixed number of gradient descent inference steps. Among these deep neural network-based approaches, MLP and TNRD have shown promising performance and can approach the denoising quality of BM3D [47]. In Chen, et al. [48] researchers proposed NAFNet, a parameter-efficient network with nonlinear activation-free layers, which outperformed state-of-the-art (SOTA) methods and demonstrated computational efficiency. Chen, et al. [49] proposed a simple GAN that takes Gaussian noise as input to generate noisy patches. However, as with most conventional methods, this GAN operates at the image level, treating images as samples and attempting to approximate the probability distribution of real-world noisy images. To overcome the limitations of this approach, a new Pixel-level Noise-aware Generative Adversarial Network (PNGAN) was introduced in [50]. This novel method performs alignment in both the image space and noise space simultaneously during training, leading to more accurate noise modeling. 1350 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate In Zhang, et al. [51] the structure of feedforward denoising convolutional neural networks (DnCNNs) was explored for image noise removal. To enhance denoising performance and accelerate the training process, residual learning and batch normalization were employed. The proposed network was successfully applied to several common image degradation tasks, including Gaussian denoising, single- image super-resolution, and JPEG artifact removal. Researchers in Zamir, et al. [52] developed an efficient Transformer-based model for image denoising and restoration, capable of handling high-resolution images. To reduce computational demands, they introduced key architectural elements such as a multi-input self-attention (SA) layer and a multi-scale hierarchical module with reduced computational complexity. 3. Restormer and Residual Learning of Dncnn Networks This section focuses on two recently developed image denoising algorithms: The Restoration Transformer (Restormer) network and the Residual Learning of the Deep CNN (DnCNN) network. Convolutional Neural Networks (CNNs), which have demonstrated strong performance in image processing applications, will be utilized in the studied algorithms due to their accessibility to large-scale datasets. 3.1. Residual Learning of the Denoising CNN (DnCNN) Network Discriminative model learning for image denoising has gained significant attention recently due to its superior denoising performance. In Zhang, et al. [51] researchers explored the construction of feedforward denoising convolutional neural networks (DnCNNs) by incorporating advancements in deep architectures, learning algorithms, and regularization techniques for image denoising. Specifically, residual learning and batch normalization are employed to both accelerate the training process and enhance denoising performance. Unlike conventional discriminative denoising models that are trained for a specific noise level (e.g., Additive White Gaussian Noise - AWGN), the DnCNN model can remove Gaussian noise with unknown noise levels (i.e., blind Gaussian denoising). Using the residual learning strategy, DnCNN indirectly estimates the clean image embedded in the hidden layers. The DnCNN Network Architecture is shown in Figure 8. 3.1.1. Residual Learning Residual learning in CNNs was initially proposed to address the degradation of performance in very deep networks. As network depth increases, training accuracy can begin to degrade. With residual learning, deep CNNs can be trained more effectively, leading to improvements in tasks such as image classification and object detection [53]. The proposed DnCNN model adopts the residual learning principle by using a single residual unit to predict the residual image (i.e., the noise), rather than directly predicting the clean image. 3.1.2. Batch Normalization Despite the simplicity and effectiveness of mini-batch stochastic gradient descent (SGD), the performance of trained CNN models can degrade due to internal covariate shift, that is, changes in the distribution of nonlinear activations during training. Batch Normalization was proposed to reduce this effect by adding a normalization step before each nonlinearity, along with a scale and shift operation. Batch normalization leads to faster training, better performance, and reduced sensitivity to weight initialization [21]. 1351 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate Figure 8. DnCNN Network Architecture. 3.2. Restoration Transformer (Restormer) Network Image restoration is the task of reconstructing a high-quality image by removing degradations (e.g., noise, blur, rain streaks) from a corrupted input. Convolutional Neural Networks (CNNs) have become the preferred choice over traditional restoration approaches due to their strong performance in learning generalizable priors from large-scale data [52]. The core operation in CNNs is convolution, which offers local connectivity and translation equivariance. While these properties provide efficiency and generalization, they also introduce two key limitations: To address these issues, the self-attention (SA) mechanism was introduced, which computes a pixel’s response based on a weighted sum of all positions in the input [54, 55]. SA has become a fundamental component in Transformer models, which are optimized for parallelization and effective representation learning [56]. Transformers have demonstrated state-of-the-art performance in natural language processing [57] and high-level computer vision tasks [58, 59]. Although SA is highly effective at capturing long-range pixel interactions, its computational complexity increases quadratically with spatial resolution, making it impractical for high-resolution image processing, common in restoration tasks. Research on adapting Transformers to image restoration is still limited [60, 61]. To reduce computational burden, some methods apply SA in small 8×8 spatial windows or divide the input image into non-overlapping 48×48 patches to perform attention independently within each patch. However, limiting the spatial scope of SA conflicts with the goal of capturing true long-range pixel relationships, especially in high-resolution images. In Zamir, et al. [52] researchers proposed an efficient Transformer model capable of handling high- resolution images for restoration tasks. To manage computational demands, they introduced a multi- head self-attention layer and a multi-scale hierarchical module, which requires fewer resources than a single-scale network [58]. A progressive training strategy was adopted to help the model learn image statistics from large datasets, enhancing contextual understanding and improving quality during inference [52]. The Restormer architecture for high-resolution image restoration is shown in Figure 9. It consists of a multi-scale hierarchical design with efficient Transformer blocks. The core modules of a Transformer block include: ● Multi-Dconv Transpose Attention (MDTA): MDTA enables spatially enriched channel-wise query-key interactions, rather than operating across spatial dimensions. ● Gated-Dconv Feed-Forward Network (GDFN): GDFN performs controlled feature transformation, allowing more effective propagation of useful information. 1352 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate Figure 9. Restoration Transformer (Restormer) network architecture Zamir, et al. [52]. 4. Experimental Results and Discussions Since noise reduction in images is an indispensable step in many practical applications, it remains a classic yet active topic in low-level vision. The goal of image denoising is to obtain a clean image x from a noisy signal y that follows the image degradation model y = x + v. A common assumption is that v is additive white Gaussian noise (AWGN) with standard deviation σ. From a Bayesian perspective, when the probability distribution is known, modeling the prior of the image plays a central role in noise removal. In this section, comparisons are made between two recently introduced algorithms for image denoising, as discussed in the previous section. Different images from the BSD68 dataset were used for the comparisons. For each resulting image, the Peak Signal-to-Noise Ratio (PSNR) was calculated. PSNR calculates the peak signal-to-noise ratio between two images in decibels. This ratio is used as a quality metric between the original image and the compressed or denoised version. The higher the PSNR value, the better the quality of the denoised image. To compute PSNR, the Mean Squared Error (MSE) is first calculated using Equation (1): 𝑀𝑆𝐸 = ∑𝑀,𝑁 [𝐼1(𝑚,𝑛)−𝐼2(𝑚,𝑛)]2 𝑀×𝑁 (1) In the above equation, M and N represent the number of rows and columns in the image matrix. Then, using Equation (2), the PSNR is calculated as follows: 𝑃𝑆𝑁𝑅 = 10 𝑙𝑜𝑔10 ( 2552 𝑀𝑆𝐸 ) (2) The PSNR results for an image from the BSD68 dataset with Gaussian noise added at a level of 50 were observed and compared using different denoising algorithms, as shown in Figure 10. Among the five algorithms compared, BM3D had the lowest performance with a PSNR of 26.21 dB, while Restormer achieved the highest PSNR value with 27.849 dB. 1353 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate Original Image Noisy Image / 14.76dB Denoised - BM3D / 26.21dB Denoised - TNRD / 26.59dB Denoised - DnCNN / 26.90dB Denoised - Restormer / 27.849 dB Figure 10. Comparison of denoising results for image 0010 from the BSD68 dataset at Gaussian noise level 50. Another comparison was carried out on the images “Cameraman,” “BSD68-10,” and “BSD68-09” with different levels of Gaussian noise (10, 15, 25, and 50) using the Restormer denoising algorithm. It is evident that the PSNR value increases inversely with the noise level. The higher the Gaussian noise, the lower the PSNR value, and vice versa. The PSNR results for the denoised images are presented in Figures 11, 12, and 13. For a better comparison, a summary of the denoising results is listed in Table 1. 1354 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate σ = 10 σ = 15 σ = 25 σ = 50 PSNR 30.50 dB PSNR 29.30 dB PSNR 27.97 dB PSNR 27.96 dB Figure 11. Gaussian image denoising on the Cameraman image using the Restormer algorithm at noise levels 10, 15, 25, and 50. σ = 10 σ = 15 σ = 25 σ = 50 1355 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate PSNR 30.65 dB PSNR 29.4 dB PSNR 28.5 dB PSNR 27.91 dB Figure 12. Gaussian image denoising on image 0010 from the BSD68 dataset using the Restormer algorithm at noise levels 10, 15, 25, and 50. σ = 10 σ = 15 σ = 25 σ = 50 PSNR 31.51 dB PSNR 30.17 dB PSNR 29.05 dB PSNR 27.93 dB Figure 13. Gaussian image denoising on image 0009 from the BSD68 dataset using the Restormer algorithm at noise levels 10, 15, 25, and 50. 1356 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate Table 1. PSNR (dB) results obtained using the Restormer algorithm at noise levels 10, 15, 25, & 50. Noise Level Cameraman BSD68-09 BSD68-10 10 30.50 31.51 30.65 15 29.30 30.17 29.4 25 27.97 29.05 28.5 50 27.96 27.93 27.91 Denoising was performed on the BSD68 dataset (consisting of 68 grayscale images) with noise levels of 15, 25, and 50 using different methods. The average PSNR (dB) results were calculated and are presented in Table 2. The best results for each noise level are highlighted in bold. The comparison shows that the DnCNN and Restormer algorithms produce results competitive with the BM3D and TNRD algorithms from the literature. While Restormer outperforms DnCNN at higher noise levels (σ = 50), DnCNN yields better performance at lower noise levels (σ = 25 and σ = 15). Table 2. PSNR (dB) results obtained using different methods on the BSD68 dataset at noise levels 15, 25, & 50. Noise level BM3D TNRsD DnCNN Restormer 15 31.08 31.42 31.46 30.71 25 28.57 28.92 29.02 28.82 50 25.62 25.97 26.10 27.95 5. Conclusions This study thoroughly investigated image denoising performance using two cutting-edge deep learning networks: the Restoration Transformer (Restormer) and the Deep CNN (DnCNN) with Residual Learning. The denoising process was rigorously applied to images from the BSD68 dataset and the classic Cameraman image, both corrupted with various levels of Gaussian noise. To quantitatively assess denoising effectiveness, Peak Signal-to-Noise Ratio (PSNR) values were calculated. These values serve as a crucial metric for measuring the fidelity between the original and denoised images. The denoising capabilities of DnCNN and Restormer were directly compared against two established algorithms in image denoising literature: BM3D and TNRD. Our findings revealed that DnCNN and Restormer consistently achieved the highest PSNR results for images subjected to Gaussian noise levels of 15, 25, and 50. These results underscore the significant potential of both deep learning architectures, particularly highlighting the superior performance of transformer architectures like Restormer in image restoration tasks. Their capabilities were shown to surpass those of traditional methodologies in terms of denoising quality. Transparency: The author confirms that the manuscript is an honest, accurate, and transparent account of the study; that no vital features of the study have been omitted; and that any discrepancies from the study as planned have been explained. This study followed all ethical practices during writing. Copyright: © 2025 by the author. This open-access article is distributed under the terms and conditions of the Creative Commons Attribution (CC BY) license (https://creativecommons.org/licenses/by/4.0/). References [1] L. Fan, F. Zhang, H. Fan, and C. Zhang, "Brief review of image denoising techniques," Visual Computing for Industry, Biomedicine and Art, vol. 2, no. 1, p. 7, 2019. https://doi.org/10.1186/s42492-019-0016-7 [2] G. Eleyan and M. S. Salman, "Convergence analysis of the mixed-norm LMS and two versions for sparse system identification," Signal, Image and Video Processing, vol. 14, no. 5, pp. 965-970, 2020. https://doi.org/10.1007/s11760- 019-01628-9 https://creativecommons.org/licenses/by/4.0/ https://doi.org/10.1186/s42492-019-0016-7 https://doi.org/10.1007/s11760-019-01628-9 https://doi.org/10.1007/s11760-019-01628-9 1357 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate [3] G. Eleyan, M. S. Salman, and C. Turan, "Two-dimensional sparse LMS for image denoising," in Twelfth International Conference on Electronics, Computer and Computation, 2015. https://doi.org/10.1109/ICECCO.2015.7416909 [4] M. S. Salman and A. Eleyan, "An efficient 2-d recursive inverse algorithm for image de-noising," in Paper presented at the Image Processing & Communications Challenges Conference, 2013. [5] G. Eleyan and M. Salman, "Image denoising with two-dimensional zero attracting LMS algorithm," Journal of Engineering Sciences, vol. 25, no. 5, pp. 539-545, 2019. https://doi.org/10.5505/pajes.2018.06982 [6] V. Atamoradov, A. Eleyan, and B. Kralik, "Performance evaluation for face recognition using wavelet-based image de- noising," presented at the International Conference on Technological Advances in Electrical, Electronics and Computer Engineering, 2013. [7] A. M. Ashir and A. Eleyan, "A multi-resolution approach for edge detection using ant colony optimization," in 23rd Signal Processing and Communications Applications Conference, 2015. [8] A. Muhammad, I. Bala, M. S. Salman, and A. Eleyan, "DWT subbands fusion using ant colony optimization for edge detection," in 22nd Signal Processing and Communications Applications Conference, 2014. [9] J. Zhang and B. Ghanem, "ISTA-Net: Interpretable optimization-inspired deep network for image compressive sensing," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2018. [10] M. Alkandari, F. Almanaye, M. Alsoori, H. Farfoura, A. Alshammari, and A. Eleyan, "Kuwaiti-Arabic sentiment analysis: A comparative study," in 6th International Conference on Bio-Engineering for Smart Technologies, 2025. [11] B. Alwazzan et al., "Brain activity monitoring using machine learning and dry-eeg signals," in Second Jordanian International Biomedical Engineering Conference, 2024. [12] M. Alkandari, F. Almanaye, S. Alhashan, D. Alenezi, and A. Eleyan, "Deep learning-based eeg signals interpretation for motor imagery," in International Conference on Bio-engineering for Smart Technologies, 2025. [13] A. Eleyan, F. Bayram, and G. Eleyan, "Spectrogram based arrhythmia classification using three-channel deep learning model with feature fusion," Applied Sciences, vol. 14, no. 21, p. 9936, 2024. https://doi.org/10.3390/app14219936 [14] Z. Lu et al., "Fast single image super-resolution via dilated residual networks," Institute of Electrical and Electronics Engineers. Access, vol. 7, pp. 109729-109738, 2018. [15] Y. LeCun, L. Bottou, Y. Bengio, and P. Haffner, "Gradient-based learning applied to document recognition," Proceedings of the Institute of Electrical and Electronics Engineers, vol. 86, no. 11, pp. 2278-2324, 2002. [16] A. Krizhevsky, I. Sutskever, and G. E. Hinton, "Imagenet classification with deep convolutional neural networks," Advances in Neural Information Processing Systems, vol. 25, 2012. [17] A. Daalah, S. Aldousari, S. Alrashed, D. Alenezi, A. Alhajri, and A. Eleyan, "Plant- leaf disease classification using fine-tuned pre-trained cnn models," in 6th International Conference on Bio-Engineering for Smart Technologies, 2025. [18] L. Bottou, "Large-scale machine learning with stochastic gradient descent," in 19th International Conference on Computational Statistics, 2010. [19] K. Simonyan and A. Zisserman, "Very deep convolutional networks for large-scale image recognition," arXiv preprint arXiv:1409.1556, 2014. [20] C. Szegedy et al., "Going deeper with convolutions," in Proceedings of the nstitute of Electrical and Electronics Engineers.History Conference on Computer Vision and Pattern Recognition, 2015. [21] K. He, X. Zhang, S. Ren, and J. Sun, "Deep residual learning for image recognition," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 2016. [22] G. Sehu, A. B. Ashir, and A. Eleyan, "Character recognition using correlation & hamming distance," in Signal Processing and Communications Applications Conference, 2015. [23] J. Larour, L. E. Aranchuk, Y. Danisman, A. Eleyan, and M. F. Yilmaz, "Modeling of the l-shell copper x-pinch plasma produced by the compact generator of ecole polytechnique using pattern recognition," Physics of Plasmas, vol. 23, no. 3, 2016. [24] F. M. Yilmaz, A. Eleyan, L. E. Aranchuk, and J. Larour, "Spectroscopic analysis of X-pinch plasma produced on the compact LC-generator of Ecole Polytechnique using artificial neural networks," High Energy Density Physics, vol. 12, pp. 1-4, 2014. https://doi.org/10.1016/j.hedp.2014.04.001 [25] G. Hu et al., "When face recognition meets with deep learning: An evaluation of convolutional neural networks for face recognition," in Proceedings of the IEEE international Conference on Computer Vision Workshops, 2015. [26] I. Mesecan, A. Eleyan, and B. Karlık, "Sift-based iris recognition using sub-segments, the international conference on technological advances in electrical," in Electronics and Computer Engineering 2013. [27] A. M. Ashir and A. Eleyan, "Compressive sensing based facial expression recognition," in 24th Signal Processing and Communication Application Conference., 2016. [28] A. Eleyan, K. Kose, and A. E. Cetin, "Image feature extraction using compressive sensing,image processing " in Communications Challenges Conference, 2013. [29] A. Eleyan, H. Demirel, and H. Ozkaramanli, "Face recognition using dual-tree wavelet transform," in Proceedings of the IEEE International Symposium on Signal Processing and Information Technology, 2008. [30] A. M. Ashir, A. Eleyan, and B. Akdemir, "Facial expression recognition with dynamic cascaded classifier," Neural Computing and Applications, vol. 32, no. 10, pp. 6295-6309, 2020. https://doi.org/10.1007/s00521-019-04138-4 https://doi.org/10.1109/ICECCO.2015.7416909 https://doi.org/10.5505/pajes.2018.06982 https://doi.org/10.3390/app14219936 https://doi.org/10.1016/j.hedp.2014.04.001 https://doi.org/10.1007/s00521-019-04138-4 1358 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate [31] A. Eleyan, "Statistical local descriptors for face recognition: A comprehensive study," Multimedia Tools and Applications, vol. 82, no. 21, pp. 32485-32504, 2023. https://doi.org/10.1007/s11042-023-14482-2 [32] M. Al-Azzeh, A. Eleyan, and H. Demirel, "PCA-based face recognition from video using super-resolution," in 23rd International Symposium on Computer and Information Sciences, 2008. [33] A. Eleyan, "Breast cancer classification using moments," in 20th Signal Processing and Communications Applications Conference, 2012. [34] A. Eleyan, Z. Al-Barakeh, R. Ghandour, B. Neji, and A. Eleyan, "Brain tumor detection via ensemble cnn-based deep learning models," in 6th International Conference on Bio-engineering for Smart Technologies, 2025. [35] A. Eleyan and E. Alboghbaish, "Multi-classifier deep learning based system for ECG classification using Fourier transform," in 5th International Conference on Bioengineering for Smart Technologies, 2023. [36] Q. Li, W. Cai, X. Wang, Y. Zhou, D. D. Feng, and M. Chen, "Medical image classification with convolutional neural network," in 13th International Conference on Control Automation Robotics & Vision, 2014. [37] E. Alboghbaish, A. Eleyan, and A. Eleyan, "Performance comparison between transform-based deep learning approaches for ecg signal classification," in 11th International Conference on Electrical and Electronics Engineering, 2024. https://doi.org/10.1109/ICEEE62185.2024.10779262 [38] A. Radford, L. Metz, and S. Chintala, "Unsupervised representation learning with deep convolutional generative adversarial networks," arXiv preprint arXiv:1511.06434, 2015. [39] L. Tran, X. Yin, and X. Liu, "Disentangled representation learning gan for pose-invariant face recognition," in Proceedings of theInstitute of Electrical and Electronics Engineers. Conference on Computer Vision and Pattern Recognition, 2017. [40] M. Kas, A. Chahi, I. Kajo, and Y. Ruichek, "EigenGAN: An svd subspace-based learning for image generation using Conditional GAN," Knowledge-Based Systems, vol. 293, p. 111691, 2024. https://doi.org/10.1016/j.knosys.2024.111691 [41] V. Jain and S. Seung, "Natural image denoising with convolutional networks," Advances in Neural Information Processing Systems, vol. 21, pp. 769–776, 2009. [42] X. Lan, S. Roth, D. Huttenlocher, and M. J. Black, "Efficient belief propagation with learned higher-order markov random fields," in Proceedings of the European Conference on Computer Vision, 2006. [43] H. C. Burger, C. J. Schuler, and S. Harmeling, "Image denoising: Can plain neural networks compete with BM3D?," in Conference on Computer Vision and Pattern Recognition 2012. [44] J. Xie, L. Xu, and E. Chen, "Image denoising and inpainting with deep neural networks: Advances in neural " in Information Processing Systems, 2012. [45] M. Elad and M. Aharon, "Image denoising via sparse and redundant representations over learned dictionaries," Institute of Electrical and Electronics Engineers. Transactions on Image processing, vol. 15, no. 12, pp. 3736-3745, 2006. [46] Y. Chen and T. Pock, "Trainable nonlinear reaction diffusion: A flexible framework for fast and effective image restoration," Institute of Electrical and Electronics Engineers. Transactions on Pattern Analysis And Machine Intelligence, vol. 39, no. 6, pp. 1256-1272, 2016. [47] K. Dabov, A. Foi, V. Katkovnik, and K. Egiazarian, "Image denoising by sparse 3-D transform-domain collaborative filtering," Institute of Electrical and Electronics Engineers.Transactions on Image Processing, vol. 16, no. 8, pp. 2080-2095, 2007. [48] L. Chen, X. Chu, X. Zhang, and J. Sun, "Simple baselines for image restoration," in Proceedings of the European Conference on Computer Vision, 2022. [49] J. Chen, J. Chen, H. Chao, and M. Yang, "Image blind denoising with generative adversarial network based noise modeling," in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018. [50] Y. Cai, X. Hu, H. Wang, Y. Zhang, H. Pfister, and D. Wei, "Learning to generate realistic noisy images via pixel- level noise-aware adversarial training," Advances in Neural Information Processing Systems, vol. 34, pp. 3259-3270, 2021. [51] K. Zhang, W. Zuo, Y. Chen, D. Meng, and L. Zhang, "Beyond a gaussian denoiser: Residual learning of deep cnn for image denoising," Institute of Electrical and Electronics Engineers. Transactions on Image Processing, vol. 26, no. 7, pp. 3142-3155, 2017. [52] S. W. Zamir, A. Arora, S. Khan, M. Hayat, F. S. Khan, and M. Yang, "Restormer: Efficient transformer for high- resolution image restoration," in Proceedings of the 2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022. [53] S. Ioffe and C. Szegedy, "Batch normalization: Accelerating deep network training by reducing internal covariate shift," in Proceedings of the International Conference on Machine Learning, 2015. [54] A. Dosovitskiy et al., "An image is worth 16x16 words: Transformers for image recognition at scale," arXiv preprint arXiv:2010.11929, 2020. [55] H. Zhang, I. Goodfellow, D. Metaxas, and A. Odena, "Self-attention generative adversarial networks," in Proceedings of the International Conference on Machine Learning, 2019. [56] S. Khan, M. Naseer, M. Hayat, S. W. Zamir, F. S. Khan, and M. Shah, "Transformers in vision: A survey," ACM Computing Surveys vol. 54, no. 10s, pp. 1-41, 2022. https://doi.org/10.1007/s11042-023-14482-2 https://doi.org/10.1109/ICEEE62185.2024.10779262 https://doi.org/10.1016/j.knosys.2024.111691 1359 Edelweiss Applied Science and Technology ISSN: 2576-8484 Vol. 9, No. 7: 1345-1359, 2025 DOI: 10.55214/25768484.v9i7.8918 © 2025 by the author; licensee Learning Gate [57] W. Fedus, B. Zoph, and N. Shazeer, "Switch transformers: Scaling to trillion parameter models with simple and efficient sparsity," Journal of Machine Learning Research, vol. 23, no. 120, pp. 1-39, 2022. https://arxiv.org/abs/2101.03961 [58] N. Carion, F. Massa, G. Synnaeve, N. Usunier, A. Kirillov, and S. Zagoruyko, "End-to-end object detection with transformers," in European Conference on Computer Vision, 2020. [59] H. Touvron, M. Cord, M. Douze, F. Massa, A. Sablayrolles, and H. Jégou, "Training data-efficient image transformers & distillation through attention," in International Conference on Machine Learning, 2021. [60] H. Chen et al., "Pre-trained image processing transformer," in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021. [61] J. Liang, J. Cao, G. Sun, K. Zhang, L. Van Gool, and R. Timofte, "Swinir: Image restoration using swin transformer," in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2021. https://arxiv.org/abs/2101.03961