3D Information from Scattering Media Images

3D Information from Scattering Media Images

scattering media, 3D depth, transmission, statistical prior, airlight

Haze, fog, and bad weather conditions occur often in daily life. In these scattering media environments,
micro-particles interfere with light propagation and image formation. Images captured in these conditions
will suffer from low contrast and loss of intensity, hindering many computer vision methods. Thus, many approaches attempt to estimate the corresponding clear scene before processing the image further. However, the image formation model in scattering media shows that the 3D distance information is encoded implicitly in image intensities. In this paper, we provide a systematic review on methods to estimate relative depth
and explicit depth directly from scattering media images. We use a dataset consisting of synthesized
hazy images with known ground truths to establish accuracy, as well as real hazy images for a general
visual analysis. For the accuracy evaluation, we demonstrate transmission estimation using statistical priors obtaining an average SSIM of 0.411 and MAE of 1.004; and depth map estimation using deep networks with an average SSIM of 0.305 and MAE of 0.860. Furthermore, for additional visual analysis, we also present the distance estimation for real hazy and underwater images of which we have no ground truth

Laksmita Rahadianti

http://dx.doi.org/10.21609/jiki.v14i1.963

Faculty of Computer Science Universitas Indonesia

2021-02-28

Sri Wahyuni

e-ISSN : 2502-9274 printed ISSN : 2088-7051

PDF

English

Text

Jurnal Ilmu Komputer dan Informasi (Journal of Computer Science and Information)