Low Light Image Enhancement With Small Training Dataset Using EnlightenGAN

Low light images have a lot of issues with visibility, low contrast, and a lot of noise. These cause difficulties in both manual human vision and systems that rely on computer vision in their decision-making. So far, various methods have been proposed to enhance low light images, ranging from traditional approaches, deep learning approaches, to adversarial learning approaches. All of these approaches must be trained on large unpaired datasets. These large datasets take a long time to collect. In addition, the training time also becomes longer due to a large number of datasets. Based on these conditions, in this study, we propose EnlightenGAN in a smaller dataset environment to reduce training and processing time but with optimal results. We tried to experiment with training using a smaller dataset, namely 100 low/normal light images, 50 low/normal light images, and 25 low/normal light images respectively. The results of the test scores obtained show that the EnlightenGAN 100 method can compete with the full EnlightenGAN dataset by each getting the highest score 5 times. Meanwhile, EnlightenGAN 50 was able to get the highest score twice and EnlightenGAN 25 got the poorest result with no highest score achieved at all. From the histogram results, it was found that the most optimal methods for equalizing histograms were EnlightenGAN 100 and EnlightenGAN 50. And by reducing this dataset, the training time is significantly reduced. From these experiments, we can conclude that the optimal minimum dataset number is above 50 training images.