碩士 / 國立中山大學 / 資訊工程學系研究所 / 107 / Single image dehazing is one of the image processing applications which many people try to realize using the advanced deep neural network (DNN) technology in recent years. This thesis proposes a dehazing network to not only dehaze the land scene images, but also to enhance the quality of underwater images. Proposed model is trained with unpaired datasets based on the generative adversarial network (GAN) approach. The GAN-based land image dehazing work has been proposed before, but this thesis has further made the following contributions. First, an AOD-net dehazing model has been integrated with the atmosphere scattering model to provide more stable inference of global atmospheric light, and better dehazing quality for objects located at far distance in the scene. Second, in addition to the ordinary reconstruction loss, the thesis also proposes several special loss functions including edge-preserving, transmission, and content losses to regularize our training process. Finally, both land and underwater dehazing processes can be unified into the same model with different parameter sets. The experimental results have shown that proposed models can outperform previous dehazing methods no matter based on traditional image processing or DNN approaches. According to the image quality metrics of PSNR and SSIM, proposed models can achieve the best metrics with small variances. The visual result of our dehazed nature images also show the proposed model can generate perceptually appealing enhanced images.
| Identifer | oai:union.ndltd.org:TW/107NSYS5392038 |
| Date | January 2019 |
| Creators | Sheng-Wei Hsu, 許勝為 |
| Contributors | Yun-Nan Chang, 張雲南 |
| Source Sets | National Digital Library of Theses and Dissertations in Taiwan |
| Language | zh-TW |
| Detected Language | English |
| Type | 學位論文 ; thesis |
| Format | 75 |
Page generated in 0.0062 seconds