visionlab task-aware image downscaling · your title here: maybe add some pictures and/or school...
TRANSCRIPT
From a single joint network of TAU and TAD
• Just recursively apply the scaling network g and f
Heewon Kim, Myungsub Choi, Bee Lim, and Kyoung Mu Lee
Department of ECE, ASRI, Seoul National University, Seoul, Korea
Task-Aware Image Downscaling
https://cv.snu.ac.kr
ComputerVisionLabSeoul National University
Problem Statement
INTRO & MOTIVATION OUR APPROACH
SR-Aware Downscaling
Normal Downscaling• Resizing an image to a smaller scale while preserving the visual appearance.• Simple downscaling operations are widely used in real world applications.• However, the inverse problems of downscaling are highly ill-posed.
Super-Resolution• Previous works on super-resolution are mostly trained with pairs (original HR image - bicubic
downscaled LR image)• Images upscaled with GAN look realistic, but can be different from the original image.• Informative low-resolution images are much easier to restore.
Image Colorization• Upscaling in the channel dimension can be applied to colorizing gray-scale images.• Adding important color information to the gray-scale image can greatly alleviate color-
ambiguity.
Loss Function
𝜃𝑓∗, 𝜃𝑔
∗ = 𝑎𝑟𝑔𝑚𝑖𝑛𝜃𝑓,𝜃𝑔1
𝑁
𝑛=1
𝑁
𝐿𝑡𝑎𝑠𝑘 𝑓𝜃𝑓 𝑔𝜃𝑔 𝐼𝑛𝐻𝑅 , 𝐼𝑛
𝐻𝑅
𝐿𝑡𝑎𝑠𝑘 = 𝐿𝑆𝑅 𝑓 𝐼𝑇𝐴𝐷 , 𝐼𝐻𝑅 + 𝝀𝐿𝑔𝑢𝑖𝑑𝑒 𝐼𝑇𝐴𝐷 , 𝐼𝑔𝑢𝑖𝑑𝑒
(We use L1 loss for both 𝐿𝑆𝑅 and 𝐿𝑔𝑢𝑖𝑑𝑒.)
Architecture & Training
Convolutional encoder-decoder model trained with a guidance image
w.r.t. the task of interest.
When we want to downscale an image:
• Just apply the downscaling network g: 𝐼𝑇𝐴𝐷 = 𝑔𝜃𝑔∗ 𝐼𝐻𝑅
When we want to upscale it later in time:
• Just apply the upscaling network f: 𝐼𝑇𝐴𝑈 = 𝑓𝜃𝑓∗ 𝐼𝑇𝐴𝐷
Upscaling a 240p TAD image to 1080p takes only 0.14s !! (on Titan Xp)
Inference
EXPERIMENTS
Colorization-Aware Downscaling
TAD preserves important information by adding task-aware knowledge at the downscaling stage.
It can be used for efficient restoration of image.
Image Upscaling (Tasks of Interest)
Downscaling vs. Compression
Deep Image Compression• Towards Image Understanding from Deep Compression without Decoding [1]
• Output of [1] : encoded bitstream only its paired decoder can convert it to an image.• Output of TAD (ours) : downscaled image can be used as a thumbnail image as-is.
Task-Aware Downscaling(TAD)• Our new downscaling approach that preserves important information w.r.t. the task of interest.• Makes the inverse problem less ill-posed and easier.
Analysis of TAD
Comparison with the SotA Super-Resolution
Extreme Super-Resolution
Only 15x12 pixels !!
More Results
Comparison to standard image compression methods (compressed output file size & reconstructed performance)
Hyper-parameter 𝜆• Control image quality trade off between HR and LR