Supervised Multi-Domain GAN for Low Light Image Synthesis
Authors
Shah, Shristy ; Kokane, Sanket Suresh
Term
2. term
Publication year
2022
Submitted on
2022-05-25
Abstract
Deep-learning methods that brighten and restore low-light photos require large amounts of training data. Today, many systems mix real low-light images with synthetic ones made by darkening normal images (for example by adjusting gamma and adding noise). However, these datasets and simple degradation pipelines often fail to capture the complex look of real low-light scenes. The RELLISUR dataset addresses this by pairing large numbers of real low-light images with their real normal-light counterparts, but it still does not cover every use case. To extend RELLISUR and provide a more realistic degradation method, we introduce a supervised multi-domain generative adversarial network (GAN) that translates normal-light images into low-light images. A GAN is a type of deep learning model that learns to generate realistic images; here, supervision and the multi-domain design let the model produce several targeted darkness levels. Our system takes real normal-light images and generates low-light versions with a specified underexposure from −2.5 EV to −4.0 EV, where EV (exposure value) measures image brightness and more negative values mean darker images. In experiments, the generated results were more varied and closer in appearance to real low-light photos than common synthetic approaches, offering a stronger basis for low-light image enhancement research.
Dybdelæringsmodeller, der skal forbedre og lysne billeder taget i svagt lys, kræver store mængder træningsdata. I dag kombinerer mange systemer ægte svaglys-billeder med syntetiske billeder, der laves ved at gøre normale billeder mørkere (fx ved at justere gamma og tilføje støj). Men både datasæt og simple forringelsesmetoder fanger ofte ikke den komplekse karakter af rigtige svaglys-situationer. RELLISUR-datasættet imødekommer dette ved at parre et stort antal ægte svaglys-billeder med deres ægte normallys-modstykker, men dækker stadig ikke alle brugsscenarier. For at udvide RELLISUR og tilbyde en mere realistisk forringelsesmetode præsenterer vi en superviseret multi-domain generative adversarial network (GAN), der oversætter normallys-billeder til svaglys-billeder. En GAN er en type dybdelæringsmodel, som lærer at skabe realistiske billeder; her gør supervisionen og multi-domain-designet det muligt at generere flere målrettede mørkhedsniveauer. Vores system tager ægte normallys-billeder og laver svaglys-versioner med en specificeret undereksponering fra −2,5 EV til −4,0 EV, hvor EV (eksponeringsværdi) måler billedets lysstyrke, og mere negative værdier betyder mørkere billeder. I forsøg gav metoden mere varierede resultater, der var tættere på udseendet af rigtige svaglys-fotos end almindelige syntetiske tilgange, og den giver dermed et stærkere grundlag for forskning i forbedring af svaglys-billeder.
[This apstract has been rewritten with the help of AI based on the project's original abstract]
Other projects by the authors
Kokane, Sanket Suresh: