DSC-Net: Single Image Haze Removal based on Depthwise Separable Convolution

Can Leng ； Gang Liu

《International Core Journal of Engineering》 9卷9期 (2023/09) Pp. 255-264

https://doi.org/10.6919/ICJE.202309_9(9).0031

摘要

Single image haze removal is an extremely difficult problem. As a hot topic in the field of computer vision, deep learning has been used to get plausible dehazing solutions. Dehaze-Net is one of the popular deep learning models proposed for single image dehazing, with excellent performance while maintaining efficiency and ease of use. However, in complex scenes, Dehaze-Net can't fully extract the boundary information. In this paper, we propose a new model called DSC-Net, which can retain the advantages of Dehaze-Net while overcoming its shortcomings to some extent. DSC-Net uses multiple groups of convolutions and maxout unit which can generate almost all haze-relevant features. And the depthwise separable convolution in DSC-Net effectively reduces the number of parameters. The numerical experimental results show that DSC-Net performs better than other similar models, e.g., Dehaze-Net, PD-Net, especially in the large sky area of the image and detail processing.

關鍵字

DSC-Net ； Image Dehazing ； Deep CNN ； Dehaze-Net ； Depthwise Separable Convolution

參考文獻

S.G. Narasimhan and S.K. Nayar:Vision and the Atmosphere, Int’l J. Computer Vision, vol. 48(2002), p. 233-254.

S.M. Pizer, E.P. Amburn, J.D. Austin, et al., "Adaptive histogram equalization and its variations", Computer vision, graphics, and image processing, vol.39(1987), p. 355-368.

J.E. McDonald:The saturation adjustment in numerical modelling of fog, Journal of the Atmospheric Sciences, vol.20(1963), p.476-478.

H.G. Adelmann,:Butterworth equations for homomorphic filtering of images, Computers in Biology and Medicine, vol.28(1998), p.169-181.

E.J. McCartney, Optics of the Atmosphere: Scattering by Molecules and Particles, vol.1(1976). New York, NY , USA: Wiley.