A lightweight filtering model for extremely dark burst image enhancement

Visual data is an important part of the modern technologies. Especially in areas such as biomedical imaging, night vision, forensic applications, underwater imaging, defense industry, smart public transportation and moving object tracking where image processing technologies are extensively used, quality of visual data is very important in order to achieve successful results, However, in cases where the ambient light is extremely low, the desired brightness level cannot be achieved in images obtained through digital photography. A digital camera sensor converts the light captured by the lens into a digital signal and stores it as a digital image. Under low levels of ambient light, the camera sensors may not be able to gather enough light to produce clear and well exposed images. This can result in dark images with increased noise and reduced image quality. Standard photography techniques and traditional enhancement methods can produce acceptable outputs for some cases but they cannot offer a satisfactory solution for images taken under extremely low ambient light. In recent years, various data-driven methods are proposed which can achieve successful results for this problem since they are able to model more complex non-linear relationships between extremely dark input and long exposure ground truth images. Inspired by these data-driven approaches, this thesis study focuses on a lightweight filtering model to process a burst set of raw input images across multiscale features for producing enhanced output images with a lower model weight and a shorter runtime. The amplification ratio scaling multiplier which is proposed as a data augmentation technique has improved the output results. The quantitative and qualitative results of the experiments show that the proposed model obtains highly competitive results with a lighter model weight and a shorter processing time when compared to state of the art studies.