Semi-Automatic Image Annotation System using Spline Modeling and Incremental Learning
- 주제(키워드) Image annotation , Image segmentation , Incremental Learning , Deep Learning
- 주제(DDC) 006.31
- 발행기관 아주대학교 일반대학원
- 지도교수 구형일
- 발행년도 2025
- 학위수여년월 2025. 2
- 학위명 석사
- 학과 및 전공 일반대학원 인공지능학과
- 실제URI http://www.dcollection.net/handler/ajou/000000034462
- 본문언어 영어
- 저작권 아주대학교 논문은 저작권에 의해 보호받습니다.
초록/요약
In this thesis, we propose a Semi-automatic Image Annotation System that uses spline modeling and incremental learning. The system is specifically designed to handle annotation tasks for images with smooth boundaries or curved shapes, such as those found in medical and industrial datasets, where traditional polygon-based methods are less efficient. Polygon annotation approximates object boundaries using linear interpolation and represents them as polygons, but it requires a large number of points to represent curved features precisely, leading to increased manual effort and inefficiency. To address these limitations, the proposed system employs spline modeling and segmentation models to reduce the manual workload and annotation time. The system generates initial annotations automatically through a segmentation model, which are then refined by human corrections. Additionally, the system incorporates incremental learning to improve annotation quality as more data is added progressively. Through a user study involving human annotators, the proposed method has demonstrated its effectiveness, achieving a high annotation quality with an IoU (Intersection over Union) of 82.51%. The annotation process also showed significant efficiency improvements, with a 60.8% reduction in the number of required clicks and a 12.8% reduction in annotation time. These results present the usability and efficiency of the semi-automatic pipeline in practical annotation scenarios, particularly for datasets with smooth boundaries and curved structures.
more목차
1 Introduction 1
2 Related Works 4
2.1 Image Annotation 4
2.2 Image Segmentation 5
2.3 Incremental Learning 6
3 Proposed Method 8
3.1 Manual mode 9
3.1.1 Initial points sampling 9
3.1.2 Spline modeling 10
3.2 Semi-Automatic mode 11
3.2.1 Segmentation 11
3.2.2 Contour-based annotation proposal 12
3.2.3 Manual correction 14
4 Experiments and Results 15
4.1 Datasets 16
4.2 User study 16
4.2.1 Evaluation Metrics 17
4.2.2 Comparison with Polygon-Manual annotation 18
4.3 Practical use of semi-automatic mode 23
4.3.1 Incremental Annotation 23
4.3.2 Application on a Downstream Task: Jaundice Diagnosis 24
4.3.3 Quantitative Results of Jaundice Diagnosis application 25
5 Conclusions 27
