Presentation av examensarbete "Real-world low-light image enhancement using Variational Autoencoders"
Olle Eriksson presenterar sitt examensarbete "Real-world low-light image enhancement using Variational Autoencoders"
Magnus Oskarsson (Matematikcentrum)
Anders Berkeman (Ericsson AB)
Niels Christian Overgaard
Low-light image enhancement is a hard task mainly due to the amount of noise and little information stored in the dark image, classical methods usually fail due to this. In this thesis project we develop a method for low-light image enhancement based on a Conditional Variational Autoencoder (CVAE) which is a deep learning model similar to the regular Autoencoder (AE). We train models on a dataset called "Seeing in the Dark" (SID), which contains paired aligned dark unprocessed RAW images and corresponding bright noise-free RGB images. This thesis is mainly focusing on building models which accept an unprocessed RAW image input and predict a (hopefully) noise-free bright RGB image, additionally we also consider alternative input data and prediction tasks. We implement and compare different neural-network architectures, and interpolation algorithms for merging overlapping image-patches. The results of the method are compared between different input-output data. The CVAE model is also compared with a model (AE) trained with only a reconstruction loss.
Regarding the different tasks, the results are showing expected results. Models with RAW input data are performing better than the model with JPG inputs, models which predict color are performing worse than the one predicting only luminance. The output images for the model predicting only grayscale images have a bit more detail and have less artifacts. The specific type of interpolation (between predicted patches) is of great importance in case the inputs are extremely noisy, otherwise the faster methods are good enough. Comparison of different types of CVAE models with the AE shows that the very type of loss-function is not of great importance for the results, but the CVAE is holding up well. The overall results are rather a function of the network architecture and the type of input data. Overall we conclude that the CVAE model can successfully be applied to the task of real-world low-light image enhancement, for visual improvement, in case the input is not too noisy, in case of an extremely noisy input the results are not as good.