Abstract

Efficient image restoration has become critical in fields such as medical imaging, surveillance, and multimedia applications, where high visual quality is imperative. Multi-frame image restoration (MFIR) leverages information from multiple correlated frames to reconstruct high-quality images, addressing challenges like noise, motion blur, and missing data. However, existing restoration methods often struggle with artifacts, loss of fine details, or computational inefficiency. This research proposes an optimized Generative Adversarial Network (GAN) framework for MFIR, focusing on enhancing the perceptual quality and structural consistency of restored images. The proposed method integrates an advanced loss function combining perceptual loss, adversarial loss, and pixel-wise mean squared error (MSE) to achieve a balance between detail preservation and global coherence. The generator employs a multi-scale feature fusion mechanism with residual connections to extract fine-grained features from input frames. The discriminator is designed to distinguish realistic textures and sharpness effectively. The framework was tested on publicly available datasets such as Vimeo-90K and Vid4, achieving a peak signal-to-noise ratio (PSNR) of 32.87 dB and a structural similarity index (SSIM) of 0.935, outperforming state-of-the-art methods by 4.5% in PSNR and 3.8% in SSIM. These improvements were observed consistently across various degradation scenarios, including Gaussian noise, motion blur, and occlusions. The proposed model also demonstrates a 15% reduction in computational complexity compared to existing GAN-based methods, making it suitable for real-time applications.

Authors

Rashmi Awasthi, B.K. Sharma
Mandsaur University, India

Keywords

Multi-Frame Image Restoration, Generative Adversarial Networks, Perceptual Quality, Image Enhancement, PSNR and SSIM Optimization