A researcher stands before an academic research poster titled "Generative Quanta Color Imaging," presented at an academic conference or symposium. The poster lists the names of its authors: Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, and Qiang Qiu from Purdue University. The poster layout includes sections on the problem statement, methodology, results, and performance evaluation. It showcases various grayscale and colorized images of both human faces and animals, illustrating the process of converting binary images into color images using neural network techniques. Detailed charts and tables present quantitative performance metrics. The poster is illuminated by overhead lights in an auditorium setting, highlighting the emphasis on image processing and artificial intelligence in the research.
Text transcribed from the image:
DUE
RSITY.
Problem
F
rization Neural
Network
e.g. Pix2Pix)
وا
eure
F2
Generative Quanta Color Imaging
Vishal Purohit, Junjie Luo, Yiheng Chi, Qi Guo, Stanley H. Chan, Qiang Qiu
Results
Burst Imag
Poor
Colorization Results
Input Binary Image
Continuously varying exposure images
Colorized Images
Reconstructed
Color Image
Our Method
Pix2Pix-DNI
Our Method
Pix2Pix-DNI
Exposure Synthesis Method
Pix2Pix-DNI
CycleGAN-DNI
SAVI21
DLOW
AtomODE-Pix2Pix (ours)
AtomODE-CycleGAN (ours)
AFHQ (512x512)
Dog
MSE (4)/RL (†)/FID (1) MSE (!)/RL (†)/FID (4) MSE ()/RL (†)/FID ()
Cat
3.75/0.9816/61.89
9.61/0.9816/207.82
41.15/46.27/178.29
2.44/0.9654/92.30
2.29/0.9997/57.88
1.15/0.9948/60.54
5.91/0.9844/95.43
16.79/0.9824/220.90
46.27/0.9804/217.98
232/0.9489/173.65
3.17/0.9994/140.23
1.18/0.9971/87.43
Wild
8.13/0.9780/39.29
9.62/0.9820/336.38
38.87/0.9019/139.47
2.75/0.9632/179.23
2.15/0.9998/79.76
1.23/0.9979/35.75
CelebA-HQ (256x256)
Male
Female
MSE (4)/RL (†)/FID (4) MSE (4)/RL (†)/FID (!)
3.74/0.9563/56.13
4.78/0.9681/39.811
9.31/0.9824/94.78
14.26/0.9810/250.61
67.26/0.9563/111.05
69.21/0.9698/261.45
1.87/0.9655/62.57
1.37/0.9798/63.31
2.15/0.9995/53.44
1.58/0.9986/77.82
3.20/0.9995/40.08
1.87/0.9994/62.59
Table 2. Colorization results on AFHQ and CelebA-HQ datasets. The input image for all these experiments is an overexposed image at
index 70 of exposure burst. All the colorizers are trained using an image corresponding to 7 = 10 in the exposure burst. At test time
we correct the overexposed input to correspond to the one used during training. In the first column of the table, we note the combination
of the exposure correction method used and colorizer is same across all the methods.
Colorization Performance Comparison
Input
Non Adaptive Pix2Pix-DNI CycleGAN-DNI
SAVI21
DLOW Our Method
Groundtruth
Image Captured using CMOS vs
Colorization
Input Correction Our method
CMOS
BIC
SIC
Groundtruth
SIC Single Image Colorization
Results on I
Colorization on Binary Images simulated
Groundtruth Our method
Colorization on Images captured using F
Groundtruth Our method 1bQIS Input