Learning to Remove Reflections from Windshield Images

Ce Wang,Boxin Shi,Lingyu Duan

DOI: https://doi.org/10.1016/j.image.2019.06.007

IF: 3.453

2019-01-01

Signal Processing Image Communication

Abstract:Reflection is a common category of contamination in images captured in the wild, particularly, undesired reflections observed on windshield images in surveillance scene bring difficulties to the analysis of surveillance contents. Existing general reflection removal methods do not exploit the specific semantic information in windshield scenes, so they cannot effectively handle the challenges of high noise and low resolution in images taken by surveillance cameras. In this paper, we propose a reflection removal framework specially designed for surveillance windshield images, with the guidance of depth of field features introduced from different levels of blurs, and constraint of specific semantic features for vehicle retrieval. Experimental results on a newly built dataset for this specific task demonstrate that our method outperforms other state-of-the-art reflection removal methods in both visual quality and quantitative metrics.