Skip to Main content Skip to Navigation
Theses

Content-aware HDR tone mapping algorithms

Abstract : The ratio between the brightest and the darkest luminance intensity in High Dynamic Range (HDR) images is larger than the rendering capability of the output media. Tone mapping operators (TMOs) compress the HDR image while preserving the perceptual cues thereby modifying the subjective aesthetic quality. Age old painting and photography techniques of manual exposure correction has inspired a lot of research for TMOs. However, unlike the manual retouching process based on semantic content of the image, TMOs in literature have mostly relied upon photographic rules or adaptation principles of human vision to aim for the 'best' aesthetic quality which is ill-posed due to its subjectivity. Our work reformulates the challenges of tone mapping by stepping into the shoes of a photographer, following the photographic principles, image statistics and their local retouching recipe to achieve the tonal adjustments. In this thesis, we present two semantic aware TMOs – a traditional SemanticTMO and a deep learning-based GSemTMO. Our novel TMOs explicitly use semantic information in the tone mapping pipeline. Our novel GSemTMO is the first instance of graph convolutional networks (GCN) being used for aesthetic image enhancement. We show that graph-based learning can leverage the spatial arrangement of semantic segments like the local masks made by experts. It creates a scene understanding based on the semantic specific image statistics a predicts a dynamic local tone mapping. Comparing our results to traditional and modern deep learning-based TMOs, we show that G-SemTMO can emulate an expert’s recipe and reach closer to reference aesthetic styles than the state-of-the-art methods.
Document type :
Theses
Complete list of metadata

https://tel.archives-ouvertes.fr/tel-03669580
Contributor : ABES STAR :  Contact
Submitted on : Monday, May 16, 2022 - 4:54:34 PM
Last modification on : Monday, May 23, 2022 - 8:27:29 AM

File

2022UPASG013_GOSWAMI_archivage...
Version validated by the jury (STAR)

Identifiers

  • HAL Id : tel-03669580, version 1

Citation

Abhishek Goswami. Content-aware HDR tone mapping algorithms. Image Processing [eess.IV]. Université Paris-Saclay, 2022. English. ⟨NNT : 2022UPASG013⟩. ⟨tel-03669580⟩

Share

Metrics

Record views

25

Files downloads

1