This master’s thesis explores the application of diffusion models to the problem of video super-resolution (VSR), a task that aims to reconstruct high-resolution video frames from low-resolution inputs. The study is situated within the broader context of deep learning advancements, particularly the recent success of diffusion models in image generation and restoration. The work is conducted at the Cyprus University of Technology, Department of Electrical Engineering, Computer Engineering, and Informatics, and supervised by Prof. Sotirios Chatzis. The thesis addresses both the theoretical underpinnings and practical implementation of diffusion-based VSR, providing a comprehensive examination of how these generative models can be leveraged to enhance video quality while maintaining temporal coherence.
Novel Application of Diffusion Models: The thesis investigates the use of diffusion models for video super-resolution, building upon their proven capabilities in image processing. By adapting these models to the video domain, the research seeks to overcome challenges unique to VSR, such as maintaining temporal consistency across frames and handling complex motion patterns.
Analysis of Temporal Consistency: A significant focus is placed on ensuring that the generated high-resolution video frames are not only visually pleasing but also temporally coherent. The thesis likely explores architectural innovations or training strategies that address the issue of flickering and motion artifacts, which are common pitfalls in VSR tasks.
Empirical Evaluation: The work includes experimental results that demonstrate the effectiveness of diffusion-based approaches compared to traditional and other deep learning-based VSR methods. The evaluation likely covers both quantitative metrics (such as PSNR and SSIM) and qualitative assessments, showcasing improvements in visual fidelity and temporal stability.
Discussion of Limitations and Future Directions: The thesis acknowledges the computational demands of diffusion models and discusses strategies for optimizing performance, such as efficient sampling or model distillation. It also outlines potential avenues for further research, including the integration of text guidance or domain adaptation for real-world video content.
The research presented in this thesis is highly relevant to the fields of computer vision and multimedia processing, particularly as high-quality video content becomes increasingly important in applications ranging from entertainment to surveillance. By demonstrating the viability of diffusion models for video super-resolution, the thesis contributes to a growing body of work that seeks to push the boundaries of what is possible with generative models. The findings have practical implications for industries that require efficient and reliable video enhancement tools, and the methodological insights can inform future developments in both academic and commercial settings. Ultimately, this work advances the state of the art in VSR and highlights the transformative potential of diffusion models in solving complex video processing challenges.