GAN-based water droplet removal
Publicerad
Författare
Typ
Examensarbete för masterexamen
Master's Thesis
Master's Thesis
Modellbyggare
Tidskriftstitel
ISSN
Volymtitel
Utgivare
Sammanfattning
This master thesis project studies how generative adversarial networks (GANs) can
perform raindrop removal and reconstruct the scenes in single images. The models
were trained by using synthetic droplet datasets, and real droplet datasets. The
real droplet datasets were from Qian et al. datasets and the Reeds datasets [1].
The synthetic droplet datasets were generated from ground-truth images from prior
datasets. The process was done by using OpenGL. Then the generated images were
evaluated by full-reference quality metrics and non-reference quality metrics, such
as SSIM and BRISQUE, then tested in object detection by a pre-trained DETR
model and evaluated by mean average precision (mAP).
After comparing the quantitative quality of the images generated by models trained
by real droplet datasets and synthetic datasets, the result showed that the real
droplet datasets yield better image quality than the synthetic datasets. In the
object detection task, though it can enhance image quality in comparison to the
degraded images, the generated images did not improve the result in this aspect.
Thus, it was concluded that the synthetic datasets need to be more realistic to be
able to reach comparable results as the real droplets. In the object detection task,
GANs generated images from the information in the latent space. As a result, there
were some objects which were corrupted, and this made the object detection model
miss classify the objects. Apparently, it may not be suitable for high precision and
safety tasks.
In the aspect of the automated evaluation system, this thesis project investigated the
design pattern for the conceptual design of the system. However, further functional
and non-functional requirements are needed to be clarified for future implementa tio
Beskrivning
Ämne/nyckelord
GANs, computer graphic, machine learning, droplet removal, image inpainting, microservice architecture