Cited time in webofscience Cited time in scopus

ABCD : Arbitrary Bitwise Coefficient for Dequantization

Title
ABCD : Arbitrary Bitwise Coefficient for Dequantization
Alternative Title
임의의 비트 단위 계수 추정을 통한 이미지 양자화 복원
Author(s)
Woo Kyoung Han
DGIST Authors
Woo Kyoung HanSunghoon ImSang Hyun Park
Advisor
임성훈
Co-Advisor(s)
Sang Hyun Park
Issued Date
2024
Awarded Date
2024-02-01
Type
Thesis
Description
Image Processing;Deep Learning;Quantization;Low Level Vision;Implicit Neural Network
Abstract
In today's digital landscape, modern displays and content creation support high bit-depth images and videos, exceeding 8 bits. However, challenges arise when compression codecs reduce bit-depth, resulting in low bit- depth (LBD) images with less than 8 bits, causing artifacts like banding and blurriness. Existing bit depth expansion (BDE) methods fall short of producing satisfactory high bit-depth (HBD) images. To address this, we propose an implicit neural function with a bit query to recover de-quantized images from arbitrarily quantized inputs, aided by a phasor estimator to utilize neighboring pixel information. Our approach outperforms previous BDE methods on natural and animated images and has practical applications, demonstrated effectively on the YouTube UGC dataset, specifically in de-banding scenarios. This research presents a promising solution to enhance image quality in situations of bit-depth reduction. Keywords: Image Dequantization, Quantization, Low-Level Vision, Deep Learning, Implicit Neural Representation|현대 디스플레이와 콘텐츠는 8비트 이상의 이미지와 비디오를 지원한다. 그러나 압축 코덱과 같은 비트 부족 상황은 낮은 비트 깊이 (Low Bit-Depth; LBD) 이미지 (<8비트)에서 밴딩과 흐릿한 아티팩트를 발생시킨다. 이전 비트 깊이 확장 (Bit Depth Expansion; BDE) 방법들은 여전히 만족스럽지 못하고 비트 깊이 (High Bit-Depth; HBD) 이미지를 생성한다. 이를 위해 우리는 임의로 양자화된 입력 이미지에서 비트 좌표를 사용하여 양자화 문제를 해결하는 함축적 신경망 표현법 (Implicit Neural Representation; INR)을 제안한다. 또한, 편각추정기 (Phasor Estimator)를 삽입하여 주변 픽셀의 편각정보를 추정하고 이를 삽입하여 모델의 성능을 상승시킨다. 제안된 방법은 자연 이미지와 애니메이션 이미지에서 이전 BDE 방법에 비해 우수한 성능을 보인다. 또한 제안된 모델을 YouTube-UGC 데이터셋에서 밴딩 제거에 적용한 결과또한 시사한다.
Table Of Contents
I. INTRODUCTION 1
1.1 Introductory Remarks 1
1.2 Bit Depth Expansion 2
1.3 Implicit Neural Representation 4
II. Problem Formulation 8
2.1 Decomposition of Quantized Signal 8
2.2 Binary Vector Space 8
2.3 Application on the Quantized Image 9
III. Method 10
3.1 Arbitrary Bitwise Coeffiients for Dequantization 10
3.2 Phasor Estimator 11
3.3 Bit Decoding 12
IV. Experiment 13
4.1 Network Detail 13
4.1.1 Encoder & Decoder 13
4.1.2 ABCD 13
4.2 Training Strategy 14
4.3 Dataset 14
4.4 Implementation detail 15

V. Results 16
5.1 Qualitative Result 22
5.2 Quantitative Result 27
5.3 Ablation Study 28
5.3.1 Network Components 28
5.3.2 Fixed-bit Training 29
5.4 Phasor Estimation 30
5.5 Debanding 32
5.6 Histogram Analysis 34
VI. Discussion 35
6.1 Phasor Estimator 35
6.2 Artifacts from Encoder 35
6.3 Flops and Memory 36
VII. Conclusion 38
VIII. Future work 39

References 41
국문요약 45
URI
http://hdl.handle.net/20.500.11750/48115

http://dgist.dcollection.net/common/orgView/200000725860
DOI
10.22677/THESIS.200000725860
Degree
Master
Department
Department of Electrical Engineering and Computer Science
Publisher
DGIST
Related Researcher
  • 임성훈 Im, Sunghoon
  • Research Interests Computer Vision; Deep Learning; Robot Vision
Files in This Item:

There are no files associated with this item.

Appears in Collections:
Department of Electrical Engineering and Computer Science Theses Master

qrcode

  • twitter
  • facebook
  • mendeley

Items in Repository are protected by copyright, with all rights reserved, unless otherwise indicated.

BROWSE