Wang, Jing (2024) Model-based hybrid variational level set method applied to object detection in grey scale images. PhD thesis, Universti Malaysia Pahang Al-Sultan Abdullah (Contributors, Thesis advisor: Liew, Siau Chuin).
|
Pdf
Model-based hybrid variational level set method applied to object detection in grey scale images.pdf - Accepted Version Download (12MB) | Preview |
Abstract
Accurate segmentation of objects from grayscale images is fundamental in computer vision with diverse applications. In medical imaging, it enables precise detection of anatomical structures, lesions, and tumors from modalities like CT, MRI, and ultrasound for diagnosis, treatment planning, and monitoring diseases. In industrial inspection, segmentation algorithms detect product defects, cracks, or anomalies for quality control and safety. However, achieving accurate segmentation in grayscale images is a challenging endeavor due to several inherent complexities. Uneven illumination, a common issue in real-world imaging scenarios, can lead to significant variations in pixel intensities, obscuring object boundaries and complicating the segmentation process. Additionally, the presence of noise, whether from sensor imperfections or environmental factors, can further degrade image quality and introduce artifacts that hinder accurate segmentation. Moreover, complex object boundaries, particularly in scenes with occlusions, shadows, or intricate shapes, pose significant challenges for traditional segmentation methods. These methods often struggle to accurately delineate the intricate contours and regions of interest, leading to under-segmentation or over-segmentation errors. To address these challenges, the proposed thesis introduces a novel model-based hybrid variational level-set method, termed VKMHLS, specifically tailored for object detection in grayscale images. VKMHLS simplifies the Local Intensity Clustering (LIC) model and introduces a novel energy functional based on the region-based pressure function, enhancing the efficiency of segmentation for low grayscale images. Furthermore, a fast numerical implementation strategy enables swift segmentation of images and estimation of the offset field, significantly improving overall computational efficiency. To robustly and accurately segment intricate object structures in challenging grayscale datasets, VKMHLS employs a multi-layer model-based level-set structure with adaptive scale operators. These operators dynamically determine the optimal number of layers and precise scale parameters, overcoming issues with local minima and enabling successful handling of severely uneven grayscale distributions. Additionally, the thesis proposes an innovative active contour model called CER, which intelligently combines elements from the Chan-Vese (CV) model and the Region-Scalable Fitting (RSF) model. CER integrates information entropy calculations and minimizes the overall energy functional, allowing successful segmentation of regions with weak edges, strong noise interference, and uneven brightness variations across grayscale images. To tackle the persistent challenge of segmenting grayscale images with both uneven characteristics and high noise levels, a hybrid level-set algorithm based on kernel metrics is introduced. This algorithm leverages an improved multi-scale mean filter to mitigate grayscale inhomogeneity while reducing the impact of scale parameter selection. Kernel measurements and local similarity metrics suppress noise influence, enhancing robustness. Furthermore, a count gradient regularization term further reduces noise impact, ensuring precise segmentation results. Comprehensive experimental evaluations demonstrate that VKMHLS accurately segments grayscale images characterized by both inhomogeneity and noise contamination, exhibiting robust performance across various noise types. These attributes make VKMHLS a highly valuable tool for tackling realworld grayscale image segmentation challenges and enabling reliable object detection.
| Item Type: | Thesis (PhD) |
|---|---|
| Additional Information: | Thesis (Doctor of Philosophy) -- Universiti Malaysia Pahang – 2024, SV: Profesor Madya Ts. Dr. Liew Siau Chuin, NO.CD : 13680 |
| Uncontrolled Keywords: | Local Intensity Clustering (LIC) |
| Subjects: | Q Science > QA Mathematics > QA75 Electronic computers. Computer science |
| Faculty/Division: | Institute of Postgraduate Studies Faculty of Computing |
| Depositing User: | Mr. Mohd Fakhrurrazi Adnan |
| Date Deposited: | 09 Jul 2025 07:50 |
| Last Modified: | 09 Jul 2025 07:50 |
| URI: | http://umpir.ump.edu.my/id/eprint/45002 |
| Download Statistic: | View Download Statistics |
Actions (login required)
 |
View Item |