| dc.description.abstract | Image Dehazing is a famous computer vision application that has been in the researcharea for the past decade. It involves reducing or removing haze/fog froman image to extract more information and make it more visually appealing overall.In this thesis, we first explain the existing methods to solve the ill-posed problemof image dehazing. We start with the prior-based techniques, which use imageprocessing to dehaze the image after which we shift to learning-based methods,which have been recently developed and are considered state-of-the-art. Followingthis, we discuss the method proposed by us and it�s results as compared toother existing state-of-the-art methods. We have proposed a two stage image dehazingmodel which utilizes two different deep learning models. The first modelis a combination of different convolutional modules like haze detector module,Dark channel prior module, feature extraction module, spatial attention module,feature fusion module and restoration module. The other model is a GAN architecture,pix2pix GAN to be specific with different generator losses. We haveobtained PSNR score of 18.11 and SSIM score of 0.6 on NH-HAZE dataset, whilewe have obtained a PSNR score of 13.79 and a SSIM score of 0.4320 on DenseHazedataset. Along with these two datasets, we have also tested our model on RESIDEdataset which also gives comparable results.These days cascading of models is quite popular to make a complex modelwhich can solve the problem with good metrics as compared to a stand-alonemodel. We have also explored the validity of this statement by comparing the resultsof a cascaded model with ours. We explore when can a cascade model benefitthe result while consuming extra computational power. Along with all these analysis,we have also experimented with different loss functions and observed thatdifferent datasets require different loss functions for better performance. | |
