Deep learning based real-time colorful night vision

Night vision systems such as Intensified Charge-Coupled Device (ICCD) provide target recognition in low-light conditions (e.g., night vision) but are inconvenient for separating and quickly analyzing images of multiple objects in a scene. Having a colorful image like daytime in night vision systems definitely improves image analysis, situational awareness of the observer, reaction time, and perceptual analysis (human vision). The aim of the study is to convert the colored, color monochrome(green) image in night vision systems into a realistic image by converting it to color in real time using a deep learning network. Cycle-GAN, one of the techniques used in image-to-image transfer, was preferred as a method. The application of the night vision field is limited to observing nature. For this purpose, a dataset of foxes caught in the wild was selected. A dataset of night vision images was needed to train the network. No such dataset was found when searched. A night vision filter was developed to carry out the thesis work. This filter produces four distinct green night vision tones from the colored image: no moonlight, quarter moonlight, half moonlight, and full moonlight. The first contribution of the study is the fox dataset generated using the four distinct moonlight night vision filters. Second, the use of deep learning networks CycleGAN to colorize color monochrome(green) night vision images is the first time implemented in this thesis. Third, augmenting the dataset using four different Moonlight filters can also be considered as a dataset augmentation technique for the night vision dataset.