A participant at an academic conference attentively examines a research poster presented by Purdue University. The poster, titled "Generative Quanta" and authored by Vishal Purohit, Junjie Luo, Yiheng Chi, and Qi, showcases a sophisticated neural network algorithm designed to tackle issues in image colorization from low-quality binary inputs. Detailed diagrams illustrate the problem of overexposure and underexposure in original monochromatic images, with their method involving exposure correction modules leading to significantly improved colorized results. Data tables shed light on performance metrics comparing their method against others, with a set of sample images highlighting visual transformations applied to various human and animal faces. The poster is displayed within a busy exhibition hall, surrounded by other attendees immersed in similar academic presentations.
Text transcribed from the image:
9
P
Is it possible to recover
a color image from a
single one-bit image?
PURDUE
UNIVERSITY
Overexposed
Problem
Scence
F
Generative Quanta C
Vishal Purohit, Junjie Luo, Yiheng Chi, Qi
Poor
Colorization Results
Input Binary Image
Results
Continuously varying expos
Binary Camera
Colorization Neural
Network
(e.g. Pix2Pix)
Exposure
Correction
Module
e
th
Underexposed
39
a
Our method
Reconstructed
Color Image
ur Method
-Coefficient Decomposition
C
cxkxk
A'
Exposure Synthesis Method
mxkxk
cxcxm
Coefficients
Atoms
03
05
Pix2Pix-DNI
CycleGAN-DNI
SAVI21
DLOW
AtomODE-Pix2Pix (ours)
AtomODE-CycleGAN (ours)
Cat
MSE ()/RL (†)/FID (4)
3.75/0.9816/61.89
AFHQ (512 x 512)
Dog
MSE (4)/RL (†)/FID (↓)
5.91/0.9844/95.43
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
16.79/0.9824/220.90
46.27/0.9804/217.98
2.32/0.9489/173.65
3.17/0.9994/140.23
1.18/0.9971/87.43
W
MSE ()/RL
8.13/0.97
9.62/0.98.
38.87/0.90
2.75/0.962
2.15/0.99
1.23/0.99
Table 2. Colorization results on AFHQ and CelebA-HQ datasets. The input ima
index=0 of exposure burst. All the colorizers are trained using an image corr
we correct the overexposed input to correspond to the one used during training. I
of the exposure correction method used and colorizer is same across all the metho
Colorization Performance Comparison
DLOW
Input Non Adaptive Pix2Pix-DNI CycleGAN-DNI
SAVI21
Our Method
Ок