1159338_Hettiarachchi,P_2021.pdf (1.58 MB)
Rain streak removal for single images using conditional generative adversarial networks
journal contributionposted on 2021-04-22, 23:27 authored by P Hettiarachchi, Rashmika NawaratneRashmika Nawaratne, Damminda AlahakoonDamminda Alahakoon, Daswin De SilvaDaswin De Silva, Naveen ChilamkurtiNaveen Chilamkurti
Rapid developments in urbanization and smart city environments have accelerated the need to deliver safe, sustainable, and effective resource utilization and service provision and have thereby enhanced the need for intelligent, real-time video surveillance. Recent advances in machine learning and deep learning have the capability to detect and localize salient objects in surveillance video streams; however, several practical issues remain unaddressed, such as diverse weather conditions, recording conditions, and motion blur. In this context, image de-raining is an important issue that has been investigated extensively in recent years to provide accurate and quality surveillance in the smart city domain. Existing deep convolutional neural networks have obtained great success in image translation and other computer vision tasks; however, image de-raining is ill posed and has not been addressed in real-time, intelligent video surveillance systems. In this work, we propose to utilize the generative capabilities of recently introduced conditional generative adversarial networks (cGANs) as an image de-raining approach. We utilize the adversarial loss in GANs that provides an additional component to the loss function, which in turn regulates the final output and helps to yield better results. Experiments on both real and synthetic data show that the proposed method outperforms most of the existing state-of-the-art models in terms of quantitative evaluations and visual appearance.
This work was supported by a La Trobe University Postgraduate Research Scholarship.
Article NumberARTN 2214
Rights StatementThe Author reserves all moral rights over the deposited text and must be credited if any re-use occurs. Documents deposited in OPAL are the Open Access versions of outputs published elsewhere. Changes resulting from the publishing process may therefore not be reflected in this document. The final published version may be obtained via the publisher’s DOI. Please note that additional copyright and access restrictions may apply to the published version.
CategoriesNo categories selected