Advisory Committee Chair
Advisory Committee Members
Dalton S Nelson
Date of Award
Degree Name by School
Doctor of Philosophy (PhD) School of Engineering
ow-brightness and image noise are common problems in computer vision that prohibit useful information extraction from video sequences. This problem is caused by insufficient illumination, low-quality cameras, external disturbances, and other factors. Dark image enhancement requires improvement of intensity and saturation; the missing data could be replaced by matching image regions using image feature detector-descriptor pairs. We proposed new metrics for evaluating image descriptors with regard to their uniqueness. Also, we developed new image enhancement techniques based on HSV color space. The idea behind the first contribution of our dissertation is to find the best descriptor that match pixels from day to night using their neighborhood. The second contribution is a generalization of the previous idea to any number of dimensions and any descriptor type. This proposed uniqueness metric allows an easier exploration of noise effects on image descriptors. The third contribution enhances nighttime images and their saturation. The two proposed enhancement techniques are saturation mapping with the least square fitting method to find the best fit for all the input saturation points. The second technique is a pixel-by-pixel neural network-based method, where we train on each pixel of input saturation and ground truth saturation from the standard LOLdatset. Experimental results on testing image sets and real-world nighttime videos show not only improvement in non-reference quality metrics NIQE, PIQE and BRISQUE, but also remarkable improvement in a subjective video quality compared to several state-of-the-art methods.
Taarji, Mohsine, "Low Illumination Video Enhancement Based on Machine Learning" (2021). All ETDs from UAB. 560.