A pipeline for lenslet light field quality enhancement

In recent years, light fields have become a major research topic and their applications span across the entire spectrum of classical image processing. Among the different methods used to capture a light field are the lenslet cameras, such as those developed by Lytro. While these cameras give a lot of freedom to the user, they also create light field views that suffer from a number of artefacts. As a result, it is common to ignore a significant subset of these views when doing high-level light field processing. We propose a pipeline to process light field views, first with an enhanced processing of RAW images to extract sub-aperture images, then a colour correction process using a recent colour transfer algorithm, and finally a denoising process using a state of the art light field denoising approach. We show that our method improves the light field quality on many levels, by reducing ghosting artefacts and noise, as well as retrieving more accurate and homogeneous colours across the sub-aperture images.

Implementation

All of our code is now available here :

RAW data decoding
Recolouring
Denoising

Additional results

Our pipeline was applied on a subset of the freely available EPFL1 and INRIA2 datasets.

Successful colour correction

Centre view (palette) Original external view ‘Centre’ recolouring ‘Prop’ recolouring ‘Prop+centre’ reco
Ankylosaurus_and_ Diplodocus_11
Bee_12
Bee_22
Color_Chart_11
ChezEdgar2
Friends_11
Fruits2
Magnets_11
Posts2
Rose2
Vespa1

Video examples

Left side : our decoding; right side : our recolouring (Bee_22).

Left side : our recolouring; right side : our denoising (Bee_22).

Left side : our decoding; right side : our recolouring (Color_Chart_11).

Left side : our recolouring; right side : our denoising (Color_Chart_11).

Epipolar images

Horizontal epipolar images.
Dansereau decoding Our pipeline (‘centre’) Our pipeline (‘prop’) Our pipeline (‘p+c’)
Bee_12
Bee_22
Color_Chart_11
Fruits2
Magnets_11
Vertical epipolar images.
Dansereau decoding Our pipeline (‘centre’) Our pipeline (‘prop’) Our pipeline (‘p+c’)
Bee_12
Bee_22
Color_Chart_11
Fruits2
Magnets_11

Fail cases

‘Centre’ recolouring scheme fail cases. First picture is the centre view (palette), the rest are fail cases (different views in the light field). (Color_Chart_11 dataset)
Examples of tiny details not being registered properly during colour correction. First picture is the centre view (palette), second is a fail case (neighbouring view of the centre one). (Posts2 dataset)
Recolouring fail cases when extreme specular effects are present. First picture is the centre view (palette), the rest are fail cases (different views in the light field). (Vespa1 dataset)

Detailed metric results

Table 1 – Average recolouring results using PSNR. Higher values are better.
Decoded Prop Centre Prop+centre
Ankylosaurus_and_Diplodocus_11 29.2905 30.4249 30.1865 30.3422
Bee_12 22.5770 24.0342 23.8479 23.9856
Bee_22 20.7177 21.7383 21.5678 21.6599
ChezEdgar2 25.8484 26.7639 26.7429 26.7728
Color_Chart_11 21.0754 21.9253 21.6420 21.9139
Friends_11 26.9947 27.3831 27.2762 27.3273
Fruits2 21.0893 21.5367 21.4133 21.4574
Magnets_11 27.8008 28.4443 28.3576 28.4102
Posts2 25.4259 29.3566 28.8164 29.2133
Table 2 – Average recolouring results using SSIM. Higher values are better.
Decoded Prop Centre Prop+centre
Ankylosaurus_and_Diplodocus_11 0.9049 0.9324 0.9300 0.9317
Bee_12 0.6550 0.7212 0.7099 0.7191
Bee_22 0.5790 0.6156 0.6101 0.6135
ChezEdgar2 0.9058 0.9152 0.9166 0.9162
Color_Chart_11 0.8023 0.8165 0.8057 0.8170
Friends_11 0.9239 0.9299 0.9290 0.9297
Fruits2 0.6317 0.6379 0.6369 0.6375
Magnets_11 0.9106 0.9346 0.9330 0.9348
Posts2 0.8054 0.8659 0.8560 0.8637
Table 3 – Average recolouring results using S-CIELab. Lower values are better.
Decoded Prop Centre Prop+centre
Ankylosaurus_and_Diplodocus_11 11.6821 7.2841 7.3169 7.2584
Bee_12 47.9358 33.7399 34.8856 33.8488
Bee_22 62.2390 58.6591 57.2721 57.9372
ChezEdgar2 29.2363 21.7885 21.0047 21.2967
Color_Chart_11 36.2238 24.2762 26.1732 24.0729
Friends_11 13.5950 11.5322 11.0613 11.2397
Fruits2 47.8697 46.9557 44.3578 45.0527
Magnets_11 15.0985 11.0375 10.7007 10.7644
Posts2 23.6692 9.1165 9.9985 9.2863
Table 4 – Average recolouring results using CID. Lower values are better.
Decoded Prop Centre Prop+centre
Ankylosaurus_and_Diplodocus_11 0.4394 0.2110 0.1993 0.2071
Bee_12 0.4277 0.2245 0.2210 0.2223
Bee_22 0.5007 0.4455 0.4433 0.4413
ChezEdgar2 0.2260 0.2253 0.2201 0.2210
Color_Chart_11 0.4059 0.3869 0.3966 0.3876
Friends_11 0.1266 0.1141 0.1128 0.1130
Fruits2 0.3110 0.3049 0.3004 0.3007
Magnets_11 0.5034 0.3056 0.3113 0.3112
Posts2 0.0222 0.0216 0.0244 0.0228
Table 5 – Average recolouring results using Histogram distance. Lower values are better.
Decoded Prop Centre Prop+centre
Ankylosaurus_and_Diplodocus_11 0.3068 0.1583 0.1591 0.1597
Bee_12 0.3419 0.2037 0.2153 0.2045
Bee_22 0.2354 0.2238 0.1841 0.2039
ChezEdgar2 0.2189 0.1225 0.1234 0.1234
Color_Chart_11 0.2274 0.1540 0.1588 0.1536
Friends_11 0.1928 0.1185 0.1170 0.1179
Fruits2 0.1930 0.1669 0.1375 0.1482
Magnets_11 0.2768 0.1545 0.1424 0.1497
Posts2 0.4981 0.2388 0.2384 0.2437
Table 6 – Average noise level before and after denoising. Lower values are better.
Decoded Prop Prop+denoising Centre Centre+denoising Prop+centre+denoising
Ankylosaurus_and_Diplodocus_11 1.268 0.830 0.010 0.847 0.013 0.816 0.009
Bee_12 3.019 2.132 0.262 2.362 0.345 2.160 0.262
Bee_22 1.994 1.704 0.197 1.971 0.214 1.775 0.196
ChezEdgar2 1.532 1.365 0.252 1.415 0.259 1.372 0.253
Color_Chart_11 1.294 1.239 0.090 1.386 0.130 1.248 0.090
Friends_11 0.584 0.563 0.110 0.570 0.105 0.586 0.113
Fruits2 1.293 1.220 0.159 1.383 0.168 1.321 0.164
Magnets_11 1.196 0.812 0.008 0.890 0.021 0.825 0.011
Posts2 1.622 0.956 0.008 1.092 0.022 0.956 0.008

Related publications

2018

Matysiak, Pierre; Grogan, Mairéad; Le Pendu, Mikaël ; Alain, Martin; Smolic, Aljosa

A Pipeline for Lenslet Light Field Quality Enhancement Conference Forthcoming

IEEE International Conference on Image Processing (ICIP 2018), Forthcoming.

Abstract | Links | BibTeX


2017

Alain, Martin; Smolic, Aljosa

Light Field Denoising by Sparse 5D Transform Domain Collaborative Filtering Inproceedings

IEEE International Workshop on Multimedia Signal Processing (MMSP 2017) - Top 10% Paper Award, 2017.

Abstract | Links | BibTeX

This entry was posted in Light-Fields and tagged by Martin Alain. Bookmark the permalink.

About Martin Alain

Martin Alain is currently a postdoctoral researcher in the V-SENSE project at the School of Computer Science and Statistics in Trinity College Dublin. He received the Master’s degree in electrical engineering from the Bordeaux Graduate School of Engineering (ENSEIRB-MATMECA), Bordeaux, France in 2012 and the PhD degree in signal processing and telecommunications from University of Rennes 1, Rennes, France in 2016. As a PhD student working in Technicolor and INRIA in Rennes, France, he explored novel image and video compression algorithms. His research interests lie at the intersection of signal and image processing, computer vision, and computer graphics. His current research topic involves light field imaging, with a focus on denoising, super-resolution, compression, scene reconstruction, and rendering.

One thought on “A pipeline for lenslet light field quality enhancement

  1. Pingback: Light Field Super-Resolution via LFBM5D Sparse Coding | V-SENSE

Comments are closed.