Deep Learning for Low-Light Image Enhancement: A U-Net Approach

The existing solutions in the field of low light image enhancement are either too complex to implement and use or introduce unwanted noise during the enhancement process. To overcome these hurdles, we have developed a straightforward solution that prioritize simplicity yet provides high accuracy, minimizing noise and maximining enhancement quality. The main motive to develop this project was to utilize Convolutional Neural Networks (CNN) to improve the clarity of a picture clicked in low-light environments by capturing fine details of input image and enhancing those details. We trained the model using ‘Mean Squared Error’ loss and the ‘Adam’ optimizer. The output images are much more understandable, making it easier to identify object’s and colour whereas also enhancing the overall quality of image. Our project showcases the capability of CNNs for image enhancement tasks, specifically within the low-light domain. The resultant model of this project can be used for day-to-day tasks which do not require complex implementations such as surveillance in residential areas. Future works may explore further improvements in this field by either introducing custom loss functions or a deeper neural network.