火炬
水准点(测量)
计算机科学
网(多面体)
人工智能
物理
反射(计算机编程)
天体物理学
地质学
几何学
数学
大地测量学
程序设计语言
作者
Allabakash Ghodesawar,Vinod Patil,Ankit Raichur,Swaroop Adrashyappanamath,Sampada Malagi,Nikhil Akalwadi,Chaitra Desai,Ramesh Ashok Tabib,Ujwala Patil,Uma Mudenagudi
标识
DOI:10.1007/978-3-031-45170-6_48
摘要
In this paper, we propose DeFlare-Net to detect, and remove flares. Typically, flares in hand-held devices are inherent due to internal reflection of light and forward scattering of lens material. The distortions due to flares limit the applications in the field of computer vision. Research challenges towards detection and removal of flare persist due to multiple occurrences of flare with varying intensities. The performance of existing flare removal methods are sensitive to the assumption of underlying physics and geometry, leading to artefacts in the deflared image. The current approaches for deflaring involve elimination of light-source implicitly, whilst removal of flare from the image leading to loss of information. Towards this, we propose DeFlare-Net for detection, and removal of flares, while retaining light-source. In this framework, we include Light Source Detection (LSD) module for detection of light-source, and Flare Removal Network (FRN) to remove the flares. Unlike state-of-the-art methods, we propose a novel loss function and call it as DeFlare loss $$L_{DeFlare}$$ . The loss $$L_{DeFlare}$$ includes flare loss $$L_{flare}$$ , light-source loss $$L_{ls} $$ , and reconstruction loss $$L_{recon}$$ towards removal of flare. We demonstrate the results of proposed methodology on benchmark datasets in comparison with SOTA techniques using appropriate quantitative metrics.
科研通智能强力驱动
Strongly Powered by AbleSci AI