In the image, a woman is seen standing near a large computer monitor, gazing intently at the screen. The monitor appears to be displaying various graphics, demonstrating a visual demonstration. She is likely observing and trying to understand the Blind Image Deblurring demo, as mentioned in the context. The woman seems to be engrossed in the content, as she stands with her hands clasped in front of her, taking in the information presented on the monitor. The setting appears to be indoors, with a neutral background, allowing the focus to remain on the woman and the monitor.
Text transcribed from the image:
KAISTS Advanced Institute of
SAMSUNG
Science and Technology
Overview
SAMSUNG ADVANCED
INSTITUTE OF TECHNOLOGY
Real-World Efficient Blind Motion Deblurring via Blur Pixel Discretization
Insoo Kim¹.2, Jae Seok Choi¹, Geonseok Seo¹, Kinam Kwon¹, Jinwoo Shin2, Hyong-Euk Lee¹
Efficient Blind Motion Deblurring Model for Practical Usage
Solving the original deblurring regression problem
Solving the blur classification problem and discrete-to-continuous
conversion (simple regression) problem
Motivation: Image Residual Error
Object Blur: Motion region vs. Non-motion region (motion type)
No Motion Region
Low Error
1Samsung Advanced Institute of Technology (SAIT), South Korea
2Korea Advanced Institute of Science and Technology (KAIST), South Korea
Real-World Efficient Motion Deblurring
Class 2
Image
TLCVPR
SEATTLE, WA JUNE 17-21, 2024
Blur Segmentaion Map
Camera Motion
"Discrete"
"Continuous"
Classes
Class 3-
Patch
Motion
Blur Image
Discretized
Image Residual Error
Residual Error
Stage 1: Blur Pixel Discretization
Stage II: Discrete-to-Continuous Conversion
[Stage 1] Blur Pixel Discretization (Classification)
Object Motion
"
Blur Segmentation Map is optimized by conducting pixel-wise sampling on
deconvolved class images. It visually aligns with image residual error
Blur
Log-T
FTEXP
Experimental Results
Ranyack
Sharp
Image
Residual Error
Blur
Segmentation
Map
Kernel
Estimator
Sampling
Object Motion Region
Bur
Latent
Sharp
Basis Kemels
High Error
Deconvolved
Class images
Deconvolved
Image
Sharp Image
Blur Image
Uniform Blur: High-frequency vs. Low-frequ
hboring pixels)
-Frequency Region
-Low Error
Bor Pixel
Discretizac
Blur Segmentation
Map
Stage 1: Classification Model
Flapjack
[Stage II] D2C conversion (Simple Regression)
Image residual error is estimated by converting blur segmentation map,
i.e., discretized image residual error into a continuous form
B
Concatenation
Methods
GMAC
High-Frequency Region
RealBlur-J
PSNRY SSIM
- High error
MPRNet 35
MIMO-UNet+
777.80
154.41
PMIMO-UN
operty
ey component
Stipficemer 349
MAXIM-35 (35)
FFTFormer [21]
GRL-8[21]
80.21
109.99
399.50
31.76 0.922
31.92 0.929
32.65 0.933
32.48 0.529
32.84 0.535
Methods
SRN-Debilur (53
MIMO-UNet-
MPRNet
132.45
2285 28
NAFN-645
63.64
Bur
Blur Phel
Discretizer
D2C
quency property
Segmentation
Map
Image
Converter Residual Emor
Deblumed Image
SepDeblum-So
34.44
22.62 0.502
32.82 0.532
32.50 0:328
32.53 8327
Restumer (4
Uliorer-8 [38
Seg Debilar-Lou
NAFNes-64 þ
SerDeblur-Sum)
Samsung EnhanceX Google Uniblur NAFNes-32+
RealBiler-R.
PSNR SSIM
39.32 0.572
41.00 0.572
38.84 0.574
38.45 0.962
41.39 0.573
41.20 0.574
38.99 0.93
39.75 0.973
SegDebitur-S
R:580ur
GMAC
1434.82
PSNR SSIMT
32.53
8.540
154.41
33.37
41856
772.00
41963
141.00
33.69
01963
86150 3338
4.966
6354
33.97
4.966
34.44
33.96
865
6268
12.56 0.534
41:23 0.975
SepDeblur-nurs)
SegDebilan-Lours)
6268
3423
6248
33.50 0.938
41.79 6.576
SegDebitur-Lo
Stage : Debiluming Model
6258
34.63
ALSTS
[Table 1] RealBlur
[Table 2] RSBlur