홍순영 학생 Single image dehazing논문 Neurocomputing에 게재
2021.12.17 11:37
본 연구실의 홍순영 학생이 연구한 “Single Image Dehazing Based on Pixel-Wise Transmission Estimation with Estimated Radiance Patches” 제목의 논문이 Neurocomputing (Impact factor : 5.719, Rank : Q1) 에 게재가 확정되었다. 본 연구는 연세대학교 전기전자공학과 통합과정 홍순영 학생(1저자)과 강문기 교수(교신저자)가 진행하였다.
Abstract
Images acquired outdoors can be affected by atmospheric conditions, such as fog and haze, and image dehazing is used to restore scene radiance in hazy images. In image dehazing, atmospheric light and transmission estimation are essential; transmission estimation is an essential step. In particular, the dark channel prior (DCP) is widely used for the transmission estimation. When using DCP-based methods, an initial transmission map is obtained through a morphological operation based on the assumption that the scene transmission in a local area is constant. However, the depth discontinuity problem cannot be avoided, and outliers are produced in the process of refining the initial transmission map. In addition, the estimation accuracy varies depending on the scene configuration due to the limitations of DCP-based methods, which simply estimate the transmission based on pixel intensity. To overcome these problems, we propose the pixel-wise transmission estimation method with estimated radiance patches (PTERP) for image dehazing. We first approximate the transmission range in the pixel location using the transmission map obtained using DCP. A patch is then set around each pixel, and several estimated radiance patches are obtained using each value belonging to the transmission range. The transmission value in the corresponding pixel location is determined using the information from the estimated radiance patches. The transmission map is then obtained by estimating the transmission value for each pixel in the entire image. With this approach, scene radiance can be restored using the determined transmission map. We performed experiments using various images, and the results demonstrated that proposed PTERP outperformed the conventional methods both quantitatively and qualitatively.
