Generative Adversarial Networks (GANs) Improving Image to Super Resolution: A System for Creating Realistic Data and Images Enhancement

Generative Adversarial Networks (GANs) Improving Image to Super Resolution: A System for Creating Realistic Data and Images Enhancement

Himanshu Vishwakarma

Department of Computer Science

Dr. D. Y. Patil Arts, Commerce & Science, College, Pimpri Pune, India.

Nachiket Andhare

Department of Computer Science,

Dr. D. Y. Patil Arts, Commerce & Science, College Pimpri Pune, India.

Abstract- Generative Adversarial Networks (GANs) , which is a revolutionize super resolution by generating photorealistic high resolution images from low quality (low resolution) images by synthetic data and adversarial training between generator and discriminator. GANs have demonstrated significantly potential across various field in AI specialist for creating High resolution images and synthetic data from real world. Advanced GANs shown a great transformation in various sectors like healthcare, finance and image processing. The generator read the low resolution images using residual blocks and sub- pixel of the images , while the decimator gives outputs as real converted high resolution images is real or fake .Super Resolution Generative Adversarial Networks(SRGANs)2017 introduced GANs to SR ( Super Resolution) and in ESRGANs tried to add discriminators and RDB blocks for text sharpening and give more better images. This can be used in various field like Medical imaging,survillance,CCTV, Satellite.GANs mainly focus on single image task for enhancing upto 2x-8x using GANs architecture of SRGANs and ESRGAN.DIV2K and generator select low resolution files and by using data augmentation ( flops and rotations) and real world scenarios try to normalise the low resolution file or images and prevent overfitting which gives the images a real look. It can save the machine learning algorithms with scarcity, imbalance and poor data variation. In this paper many we try to provided various GANs and their implementation in real life for making life easier like image to image synthesis, style transformation.

Keywords : Generative Adversarial Networks (GANs),Super Resolution (SR),SRGANs,ESRGANs,Photorealistic resolution images, Low- resolution images,Generator,Discriminator ,Residual blocks, Subpixel convolution Adversarial training , Medical imaging, Surveillance,CCTV,Satelliteimaging,DIV2K dataset, Data augmentation, enhancement Synthetic data

1.INTRODUCTION:

Generative Adversarial Networks (GANs) are generative models in which two networks a generator and a discriminator are trained simultaneously in an adversarial framework. The generator learns to produce synthetic data from random input , while the discriminator try to distinguish these generated samples from real data., the discriminator is updated to correctly classify real vs. fake images, and. the generator gradually improves its outputs until it produces highly realistic samples that closely match the real data distribution. In short, GANs enable high-quality data synthesis by the generator and discriminator contest each others improvements.For the following topic we required various data from different sources of datasets, Synthetic data generated by GANs augments limited datasets, preserves privacy by masking sensitive information, and balances skewed distributions for machine learning. In practice, it supports diverse uses like data

augmentation and cross-domain transformations, outperforming traditional methods.

1.2.Historical Data:

At the earlier time the concept of image super resolutions is imagined in the year of 1980s and in the 2003 it made a improvement into the concept of low resolution images to High resolution and further it was developed to ERGANs when it was introduced in 2017 to SRGANs which tries to add descriminators and RDB Blocks for image sharpening and giving a realistic look of clearified images.Image super- resolution (SR) methods relied on interpolation techniques such as bilinear and bicubic interpolation, which often produced blurred images and lacked high-frequency details. the early 2000s, single-image SR emerged via example-based approaches. Freeman et al. (2003) introduced patch-matching from LR-HR training pairs, marking a shift to data-driven

1.3 Objective and scope :

priors over handcrafted rules. This era saw high level coding but the still outputs remained blurry due to reliance on pixels losses favoring averages over sharp textures. Traditional learning-based approaches later attempted to model LRHR images mappings using various features but showed limited generalization.The introduction of Generative Adversarial Networks (GANs) by Good fellow et al. in 2014 gives a important turn in image generation research. At the current time the SR models achieved high PSNR values but generated overly smooth results due to pixel-wise loss optimization.

The primary objective of this research is to explore how Generative AI focus on GANs, generates realistic synthetic data to overcome data scarcity and privacy issues in machine learning applications. it aims to enhance image super- resolution by developing GAN-based models that produce high-fidelity improved images, improving visual quality.

The main objective of the following data is to enhancement into survillence footage and all other footage,restoration of historical data or the degraded image ,synthetic dataset generation for privacy preserving research.

2.METHODOLOGY

The Scope of the GANs is not many systems but you can into the all other systems for data augmentation and generation of high quality images,In computer systems for generation of 2-D images generation and enhancement.In medical imaging field like scans MRI ,Architecture models, sattetiite monitoring and the low quality of CCTV footage to High Quality footage for surveillance and the time for training and all other factors is decreased.

jointly optimize data realism, perceptual quality, and structural consistency through a multi-stage intelligent datasets.

2.2 GENERATOR

The GAN architecture consists of a generator and a discriminator trained in an adversarial manner. The generator convert low-resolution images into high-resolution outputs by analysingthe image data missing details and textures, while the discriminator evaluates the realism of the generated images by comparing them with real high-resolution images. A combination of adversarial loss and content-based loss is used to balance perceptual realism and structural accuracy. When the generator is set to create realistic synthetic high-resolution images from low-resolution inputs. These synthetic images are added to the dataset to increase data diversity and reduce overfitting. The system is evaluated using structural similarity and visual quality measures to verify improvements in both synthetic data realism and image super-resolution performance.

2.1 SYSTEM ARCHITECTURE

Basically the generator is a heart of the GANs ,if generator stoped working then all the process will stoped just like human body.The generator in a GAN-based super-resolution system functions as an intelligent image reconstruction network that convert low-resolution images into high-resolution outputs by finding missing details. It first receives a low-resolution image as input and passes it through convolutional layers to extract basic features such as edges, textures, and colour patterns. These features are then processed through deep residual blocks, where the generator learns complex structures and high- frequency details commonly present in real high-resolution images. Residual connections help the network focus on restoringlost information rather than relearning the entire image content. After all this process, the generator increases the image resolution using upsampling techniques such as sub- pixel convolution, transforming the low-resolution feature maps into high-resolution representations. Finally, additional convolution layers refine the output by sharpening edges and enhancing textures. During training, the discriminator evaluates the generated image against real high-resolution images and provides feedback, enabling the generator to iteratively improve realism and visual quality.

2.3 Discriminator

The following research paper provide the information about GANs for clearfication of the image from low resolution to high resolution using different type of software like SRGAN and ERGAN to gives the image a realistic vibes. SRGANs methodology provides a robust framework for using GANs in super resolution image and Data creation(synthetic).The core of this idea is generator network that upsamples images and a discriminitor which check the result the given following program or real image is same as real world,It is trained for producing realistic outputs.

The system architecture is designed to Generative Artificial Intelligence and Generative Adversarial Networks (GANs) to generate realistic synthetic image data and improve single- image super-resolution. Unlike other various architectures that focus solely on pixel reconstruction, this system is structured to

Generative Adversarial Network (GAN), the discriminator functions as an intelligent decision-making network whose primary role is to evaluate the authenticity of images by distinguishing real samples from generated ones. When an imageeither real high-resolution data from the dataset or a synthetic image produced by the generatoris fed into the

2.4 Data Collection & Improvement using AI:

discriminator, it first passes through a series of convolutional layers that extract hierarchical features such as edges, textures, contrast patterns, and spatial relationships. As the image moves deeper into the network, the discriminator analyzes increasingly complex visual characteristics, including high- frequency details and global structural consistency. Through nonlinear activations and progressive downsampling, it learns to detect subtle artifacts, unnatural textures, or inconsistencies that may not be noticeable at the pixel level. The discriminator is trained using supervised signals where real images are labeled as authentic and generated images as fake, allowing it to continuously refine its decision boundary. Its output is a probability score representing how realistic the input image appears. During adversarial training, the discriminators feedback is used to update the generator, penalizing unrealistic outputs and encouraging the synthesis of more convincing images. As training progresses, the discriminator becomes more sensitive to fine visual details, forcing the generator to improve texture reconstruction, edge sharpness, and perceptual realism, thereby playing a critical role in achieving high-quality image super-resolution and realistic synthetic data generation. There is some loss function for decriminitor, Adversial losss, Content Loss, perceptual loss. Discrimintaor is work as a analyst into GANs,by evaluating image: Extarcting features(edges,image sharpening,textures and patterns), Analyzing Complex,visual characteristics and the most important outputting a probability score for realism of images.

This methodology focuses on integrating artificial intelligence mechanisms within Generative Adversarial Networks to enable intelligent data collection, refinement, and continuous improvement for image super-resolution tasks. Instead of relying on static datasets, the proposed approach uses Al-driven feedback loops to identify data scarcity, visual complexity, and resolution gaps within the training samples. The GAN model dynamically generates high-quality synthetic data that mimics real-world image distributions while adapting to domain- specific features. An Al-based evaluation module continuously analyzes generated outputs to determine their contribution to model learning, selectively retaining samples that enhance perceptual quality and structural consistency. Through iterative adversarial learning, the system improves both data diversity and super-resolution performance without external manual intervention. This adaptive methodology allows GANS to function not only as im enhancers but also as intelligent data collector and improvement of image enhancement by collecting the data which used to process the real time solution and creating super resolution footage & maintain the realation between the image obtained and the realistic manner by providing the data to the discriminator and generators.

2.5 Algorithm:

Step 1: Data Preparation Collect high-resolution (HR) images from dataset (e.g., DIV2K).Generate corresponding low- resolution (LR) images using downsampling (bicubic interpolation).

Apply data augmentation:

Rotation Flipping

Step 2: Generator Network (G):

Convolutional Layer

Residual Blocks (RDB blocks in ESRGAN)

Step 3: Generator Forward Pass Input LR image to Generator Generate Super-Resolved image

Step 4: Feed real HR image Label = 1 Feed generated SR image Label = 0 Compute Discriminator Loss

Step 5: Continue adversarial training until:

Loss stabilizes

PSNR / SSIM improves

Visual quality becomes realistic

Step 6: Input new Low-Resolution image.

Generator produces Super-Resolved image

figure1: Synthetic image generation using GANs

Figure2: The overall diagram for new GANs for medical images super resolution

3.1Result: Experimental Setup

The proposed GAN-based super-resolution system was trained using standard high-resolution image datasets such as DIV2K. Low-resolution (LR) images were generated from high- resolution (HR) images using upscaling (2×, 4×, and 8× scaling factors). Which help to give a realistic data as real world by comparing noise image which was generated using generator and real Data augmentation techniques such as flipping and rotation were applied to increase dataset diversity and prevent overfitting. The model architecture consisted of: Generator with residual blocks and sub-pixel convolution layers.Discriminator with deep convolutional layers.Loss functions: Adversarial Loss, Content Loss, and Perceptual Loss.The generated high-resolution images showed:Improved edge sharpness,Enhanced fine textures (hair, fabric, ),Reduced blurring artifacts .

In surveillance and CCTV footage enhancement: Text and number plates became more readable,Background noise was reduced In medical imaging: MRI and scan images showed improved clarity,Structural details were more visible

GAN-generated synthetic data helped: Increase dataset size without collecting new real-world data,Reduce overfitting,Improve model generalization. For medical imaging process we can get more clearified image for CT scan and MRI. For satellite monitoring process the view of map and view the places where map view ae not much clear, it can be gives more better view which help in navigation process.

4 FUTURE SCOPE:

Generative Adversarial Networks (GANs) for image super- resolution have shown significant improvements in perceptual quality and synthetic data generation. However, there are several promising directions for future research and development.

4.1 Real-Time Super-Resolution Systems

Future research can focus on developing lightweight and efficient GAN architectures that enable real-time super- resolution for:

Live CCTV monitoring ,Video streaming platforms, Real-time satellite surveillance.

5 CONCLUSION:

GANs have revolutionized image super-resolution by generating ealistic synthetic high-resolution data from low- resolution inputs, achieving superior perceptual quality and detail reconstruction. This approach outperforms traditional methods in metrics like PSNR and SSIM while addressing challenges like data scarcity.GAN-based models, such as SRGAN and ESRGAN, excel in producing visually appealing images indistinguishable from real high-resolution ones through adversarial training between generator and discriminator networks Custom loss functions balancing fidelity and perceptual metrics, along with techniques like attention mechanisms and progressive training, enhance stability and performance.Ongoing research should focus on stabilizing GAN training, improving generalization across domains like medical and satellite imaging, and integrating advanced architectures for even higher fidelity. These advancements promise broader applications in computer vision, paving the way for ethical, efficient synthetic data generation.

6. REFRENCE:

1. I. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville and Y. Bengio,

2. Generative Adversarial Nets, in Advances in Neural Information

   Processing Systems (NeurIPS), vol. 27, 2014, pp. 26722680.

3. C. Ledig, L. Theis, F. Huszár, J. Caballero, A. Cunningham, A. Acosta,

   A. Aitken, A. Tejani, J. Totz, Z. Wang and W. Shi,Photo-Realistic Single Image Super-Resolution Using a Generative Adversarial Network, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), 2017, pp. 46814690

4. W. T. Freeman, T. R. Jones and E. C. Pasztor,Example-Based Super- Resolution, IEEE Computer Graphics and Applications, vol. 22, no. 2, pp. 5665, Mar.Apr. 2002

5. E. Agustsson and R. Timofte,NTIRE 2017 Challenge on Single Image Super-Resolution: Dataset and Study, in Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW), 2017, pp. 126135.