On-orbit spatial image characterisation and restoration based on stochastic characteristic targets / Wong Soo Mee
While the qualities associated with the radiometric and geometric resolution are a major concern in earth observation satellite (EOS) imaging sensors calibration and validation, spatial resolution quality is an important parameter that is essentially needed for on-orbit spatial EOS imaging performan...
Saved in:
| Main Author: | |
|---|---|
| Format: | Thesis |
| Published: |
2021
|
| Subjects: | |
| Online Access: | http://studentsrepo.um.edu.my/14521/1/Wong_Soo_Mee.pdf http://studentsrepo.um.edu.my/14521/2/Wong_Soo_Mee.pdf http://studentsrepo.um.edu.my/14521/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | While the qualities associated with the radiometric and geometric resolution are a major concern in earth observation satellite (EOS) imaging sensors calibration and validation, spatial resolution quality is an important parameter that is essentially needed for on-orbit spatial EOS imaging performance assessment. Moreover, its calibration results can be applied to an image restoration problem, to improve the spatial quality of the EOS data. A practical way to characterize the on-orbit spatial quality performance of an EOS imaging sensor is to determine the modulation transfer function (MTF) from its remotely sensed images on the ground. However, existing approaches and techniques for spatial characterisation are highly reliant on the presence and manual identification of a well-separated characteristics target. These approaches and techniques impose stringent criteria and temporal sampling issues. In spatial image restoration, even with the perfect estimation of degradation function, restoring coherent high-frequency image details can still be very difficult.
This thesis presents two frameworks; the first one is for on-orbit spatial characterisation, whereas the second is for optical satellite image restoration. In the first framework, this thesis introduces an insight to effectively measure the MTF by analyzing the stochastic characteristics in the observed image. In particular, first, it proposes a segmentation method to select the ideal candidates for MTF Measurement. Second, it develops an adaptive structure selection method that removes detrimental structures and selects only useful information for point spread function (PSF) estimation. Finally, it introduces a spatial prior that can simultaneously suppress noises while preserving the sparsity and continuity of data to obtain high fidelity two-dimensional PSF model for MTF measurement. The experimental results demonstrate that the proposed framework is practical and effective, with < 2.3% of relative error at the Nyquist frequency as compared to the well-established edge method.
In continuation of the first framework, the proposed MTF measurement algorithms are evaluated experimentally as a blur kernel estimation method for spatially varying and invariant blur removal. Furthermore, this thesis presents a comparative study on blur estimation methods that utilize the principle of sparse representation to gain an insight into image priors type that appropriate for blur removal in blind optical satellite images. Given the fact that the heavy-tailed properties of MTF typically introduce noise and an unacceptable aliasing effect. Therefore, in the second framework, this thesis exploits the image properties and shows that only one image property used in a regularization-based framework is insufficient to obtain satisfying restoration results. Hence, this thesis presents a strategy for high-fidelity MTF compensation by characterizing both the local smooth and nonlocal self-similarity properties of images in the hybrid domain. To minimize computational complexity, it establishes a simple joint statistical model in the Curvelet domain to combine these image properties and employ the multi-objective bilevel optimization approach to efficiently solve the severely ill-posed inverse problem of MTFC. The experimental results show that the proposed methods achieve significant performance in preserving high fidelity images with feature similarity (FSIM) index value as high as 0.99876 and minimum computational complexity.
|
|---|