Enhanced Sparse Keypoint Detection in Near-Uniform Scenes for Robust Image Stitching
Image stitching is a vital image processing technique in many modern multimedia applications, particularly for creating panoramic images and virtual reality environments. However, this processing presents significant challenges in near-uniform scenes, where theoretical feature detectors often strugg...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Language: | en en en |
| Published: |
Universiti Malaysia Sarawak (UNIMAS)
2025
|
| Subjects: | |
| Online Access: | http://ir.unimas.my/id/eprint/49561/4/dsva_Jong%20Tze%20Kian.pdf http://ir.unimas.my/id/eprint/49561/5/Thesis%20PhD_Jong%20Tze%20Kian.pdf http://ir.unimas.my/id/eprint/49561/6/Thesis%20PhD_Jong%20Tze%20Kian_24%20pages.pdf http://ir.unimas.my/id/eprint/49561/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Image stitching is a vital image processing technique in many modern multimedia applications, particularly for creating panoramic images and virtual reality environments. However, this processing presents significant challenges in near-uniform scenes, where theoretical feature detectors often struggle to identify distinctive interest points. This problem arises when images contain homogeneous content that lacks salient features, making it difficult to generate a sufficient number of well-distributed matching keypoints. The absence of such features often leads to an inadequate number of matched inliers between overlapping images, resulting in undesirable visual effects such as seams, ghosting, and geometric distortions. To address these limitations, this thesis proposes two novel feature detection approaches aimed at improving image stitching performance in near-uniform scenes. The first method, referred to as the enhanced partial differential equation (ePDE), introduces a feature detection technique based on nonlinear diffusion. This approach applies the Lorentz factor, inspired by Einstein’s theory of special relativity, to modify the conductivity function within the partial differential equation. This enhancement enables the construction of multiple enhanced edge-preserving nonlinear scale-spaces, significantly improve feature detection in homogeneous regions and outperforming several state-of-the-art detectors. The second method employs the Log-Gabor feature transform (LGFT), which generates multiple oriented and scaled Log-Gabor scale-spaces to enhance the extraction of distinctive and matchable interest points, thereby addressing the limitations of intensity-based feature detection techniques. Additionally, two new evaluation metrics are introduced: the spread-overlap measure (So) and the recall-to-spread-overlap ratio (RC/So). These metrics offer a more comprehensive performance evaluation by considering both the accuracy and spatial distribution of inliers. Experimental results show that the proposed LGFT and ePDE methods consistently improve image stitching performance, especially for near-uniform image datasets. Both methods achieve over a 14.63% higher image stitching success rate compared to current state-of-the-art techniques. The top-performing LGFT-1.4.3 variant achieves an average improvement of 4.8% in inlier accuracy and distribution, while the ePDE-g5 variant records an average improvement of 4.3%, both outperforming existing feature detectors. These results highlight the robustness and effectiveness of the proposed methods in addressing the challenges of stitching near-uniform scenes. |
|---|