Research Projects

Image enhancement is essential in consumer electronics, surveillance, medical imaging, and satellite imaging, where high-quality images must be generated from degraded inputs caused by noise, blur, or low resolution. Traditional methods, especially convolutional neural networks, achieve fast enhancement but often fail to deliver high-quality results based on low-quality inputs. The recent advancements in generative methods, particularly diffusion models, demonstrate superior image enhancements by capturing complex pixel relations. However, achieving real-time performance with diffusion models while maintaining quality remains challenging. This limits real-world applications for diffusion models since use cases like live video surveillance, medical imaging, and streaming require time-efficient and high-quality models. Scientific Objectives: This project aims to create a framework for real-time image enhancement using an optimized generative diffusion model. The proposed model will operate through novel Stochastic Differential Equations (SDEs), resulting in an effective and novel method for integrating degraded image data into the process. This method combines high fidelity and fast enhancement, achieving state-of-the-art results on 1280x720 images within 60 to 100 milliseconds, irrespective of the input image quality. A key goal is to implement the algorithm on edge devices, potentially starting as an app for smartphones, to make advanced image enhancement accessible directly on portable, budget-friendly hardware. Research Hypothesis and Model: We hypothesize that conditioning the diffusion model on degraded input images through novel SDEs will enable high-quality, real-time enhancement. We will optimize reverse SDE in diffusion models, with efficient neural network parameterization, to validate this hypothesis, thereby obtaining state-of-the-art output quality under real-time constraints. Key Experiments: 1) SDE Conditioning: Test the integration of degraded data within developed novel SDEs to evaluate its impact on enhancement quality. 2) Time Efficiency: Develop techniques to reduce the number of iterations to solve the proposed differential equation efficiently. 3) Real-Time Performance: Benchmark the proposed methods against state-of-the-art image enhancement algorithms, with respect to computational time and quality on edge devices. Significance to the Field: This project offers significant technological and societal impact, particularly relevant in the Indian context. By developing real-time generative models for high-quality image enhancement, we aim to make advanced imaging accessible on budget hardware (where input images are of low quality), including smartphones, TVs, and surveillance systems. This will facilitate portable and on-demand image enhancement, enhancing public safety, healthcare, and education by providing accessible, high-quality imaging solutions at an affordable cost.