Laser & Optoelectronics Progress, Volume. 61, Issue 18, 1800004(2024)
Progress in Research and Application of Image Stitching Technology Based on Regional Optimization
Figures & Tables(19)
Fig. 1. Schematic diagram of solving the optimal seamline based on graph cuts. (a) Original images and the solved optimal seamline; (b) optimal seamline obtained using graph cuts
Fig. 3. Comparison between the traditional image stitching process and seam-driven process[23]
Fig. 6. Diagram of remote sensing image stitching based on distribution measurement and saliency information[46]
Fig. 12. Diagram of remote sensing image stitching method based on Delaunay triangular mesh[55]
Fig. 14. Schematic of pavements remote sensing image stitching segmentation detection method[68]
Table 1. Comparison of typical local adaptive warping image stitching methods
View tableTable 1. Comparison of typical local adaptive warping image stitching methods
Author Principle Advantage Disadvantage Time consuming Parallax Gao et al.［ 11 ］Dual-homography warping Align background and foreground separately Not for complex scenarios ++ No Lin et al.［ 15 ］Smoothly varying affine Stronger local deformation and alignment capability Local artifact ++ No Zaragoza et al.［ 12 ］Dense mesh High-precision local alignment Projection error +++ No Chang et al.［ 16 ］Shape-preserving half-projective Combine projective and similar transformations Prone to ghosting + No Lin et al.［ 13 ］Combination of local homography and global similarity transformations Reduce perspective distortion in non-overlapping areas Local distortion +++ No Chen et al.［ 17 ］Natural image stitching with the global similarity prior More natural stitched images Local distortion ++ Yes Li et al.［ 18 ］Robust elastic warping Elastic warping to eliminate parallax Local deformation + Yes Jia et al.［ 22 ］Leverage line-point consistence to preserve structures Preserve both local and global geometric structures for wide parallax images Complex calculation ++ Yes
Table 2. Comparison of typical optimal seamline image stitching methods
View tableTable 2. Comparison of typical optimal seamline image stitching methods
Author Principle Advantage Disadvantage Time
consuming
Parallax Zhang et al.［ 24 ］Parallax-tolerant Combine global transformation and content-preserving warping Low efficiency +++ Yes Lin et al.［ 25 ］Seamline-guided local alignment Leverage estimated seamline to guide local alignment It may fail when the parallax is too large or when there are few feature matching points ++ Yes Liao et al.［ 34 ］Quality evaluation-based iterative seamline estimation Iterative seamline estimation with seamline quality evaluation Large parallax or moving objects may fail +++ Yes Li et al.［ 35 ］Perceptual-based seamline-cutting Integrate human perception into the energy minimization process Seamline tend to appear at overlapping boundaries ++ Yes Zhao et al.［ 46 ］Outlier removal，optimal seamline detection and smooth transition Combine multiple optimization algorithms Low efficiency +++ Yes
Table 3. Comparison of typical deep learning image stitching methods
View tableTable 3. Comparison of typical deep learning image stitching methods
Author Year Advantage Disadvantage Stitching method Supervision Code Nie et al.［ 47 ］2020 View-free It suffers from problems such as the need for more accurate initialization and more computational resources Non-seamline Yes Yes Nie et al.［ 48 ］2021 Unsupervised deep image stitching Performance to be improved Non-seamline No Yes Li et al.［ 49 ］2021 Alignment of a set of matched dominant semantic planar regions Quality of matched semantic planar regions depends on segmentation results Seamline Yes No Nie et al.［ 50 ］2021 Stitching any viewing angle and size Performance can be affected by feature points Non-seamline Yes No Jiang et al.［ 51 ］2022 Progressive equivalence constraint homography estimation It still has its limitation of being applied to scenes with multiple planes Non-seamline Yes Yes Nie et al.［ 52 ］2022 First deep learning based rectangling solution Warping-based method could introduce artifacts and noise due to the lack of accuracy of warping motion fields Seamline Yes Yes Zhou et al.［ 53 ］2024 First diffusion-based learning algorithm to tackle the rectangling for image stitching For two-image stitching only Seamline Yes Yes
Table 4. Application of image stitching technology in typical fields
View tableTable 4. Application of image stitching technology in typical fields
Fields of application Requirement Characteristic Common method Challenge Issue Remote sensing Earth sciences，urbanization，environmental and resource management Large amount of data Airborne hyperspectral images，infrared images Complexity of the environment Difficulty of combining accuracy and efficiency Agriculture Obtain accurate data on farmland and crops Wide range Images taken by UAV Complex backgrounds Incomplete or missing contours Inspection Obtain accurate data for defect detection High inspection accuracy and efficiency Vehicle-mounted continuous capture of defect images Lack of rich texture information Matching error and stitching distortion Medical imaging Large-field，high-resolution images Real-time and efficient Microscopy or ultrasound Images with greater variability and complexity Difficulty with image alignment Industrial measurement Measurement and analysis of large-sized or complex structure High requirements for feature point extraction Construction of measurement devices Poor feature clarity，low contrast，weak texture Insufficient feature extraction or incorrect matching
Tools
Get Citation
Copy Citation Text
Weidong Pan, Anhu Li, Xingsheng Liu. Progress in Research and Application of Image Stitching Technology Based on Regional Optimization[J]. Laser & Optoelectronics Progress, 2024, 61(18): 1800004
Paper Information
Category: Reviews
Received: Jun. 3, 2024
Accepted: Jul. 29, 2024
Published Online: Sep. 14, 2024
The Author Email: Anhu Li (lah@tongji.edu.cn)
CSTR:32186.14.LOP241417
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20