Refocus: Transforming Low Light Imaging - Real Time Deblurring and Enhancement Using Convolutional Neural Networks

Abstract

Blurred images are a problem almost everyone encounters, whether in everyday photography or in professional fields like security, technology, and data analysis. Issues such as camera shake, slow shutter speeds, poor focus, and especially low light conditions can cause important visual information to be lost. This project focuses on addressing that challenge by developing Refocus, a deep learning based system designed to restore clarity in images that suffer from blurriness, with a special emphasis on images captured in low-light environments.

The core of this solution is a Convolutional Neural Network (CNN) trained on pairs of blurred and sharp images to learn how to reconstruct clearer versions of degraded inputs. To support this process, several preprocessing steps—such as resizing, cropping, and normalization were applied to ensure consistent and high quality training data. While existing deblurring methods show promise, many still struggle to deliver accurate results under poor lighting. This research identifies that limitation and aims to produce sharper, more detailed outputs in such challenging scenarios.

The initial prototype demonstrated encouraging results, showing the model’s ability to reduce blur and enhance image clarity. At the same time, it revealed opportunities for improvement, such as refining the architecture and exploring more advanced datasets. Overall, this work contributes to ongoing efforts in low light image restoration and lays the foundation for future enhancements in real time image deblurring using deep learning.