Training CNNs in Presence of JPEG Compression: Multimedia Forensics vs Computer Vision 论文阅读
date
Oct 3, 2023
Last edited time
Oct 3, 2023 06:32 AM
status
Published
slug
Training CNNs in Presence of JPEG Compression: Multimedia Forensics vs Computer Vision 论文阅读
tags
DL
summary
type
Post
Field
Plat
Introduce
In this work, we focus on the effect that JPEG has on CNN training considering different computer vision and forensic image classification problems. Specifically, we consider the issues that rise from JPEG compression and misalignment of the JPEG grid. We show that it is necessary to consider these effects when generating a training dataset in order to properly train a forensic detector not losing generalization capability, whereas it is almost possible to ignore these effects for computer vision tasks.
CASE STUDIES
- JPEG Grid Misalignment
神奇的问题，但是简而言之，不用管
While editing a photograph or simply uploading a profile picture over social networks, it often occurs that images are cropped with respect to their original size. As a matter of fact, this operation is performed most of the times without paying attention to the precise pixel coordinates of the cropped area, as it is more important to prevent the picture subjects being canceled by the cropping. As a consequence, it generally happens that JPEG-compressed images are cropped without respecting the 8 × 8 characteristic pixel grid introduced by JPEG compression. If the image is then further saved as JPEG, a new 8×8 grid non-aligned with the original one is generated.
- Quality factor of JPEG Compression
没意思
