周读论文系列笔记(4)-reivew-Image registration methods: a survey
文章目录
-
- 1. Introduction
-
Image registration methodology: This section delves into the technical aspects of aligning and matching images through systematic methodologies.
-
Step1: Feature detection: The initial phase involves the identification and extraction of distinct features within an image.
-
- 3.1 Area-based methods: These techniques focus on analyzing and comparing regions or areas within the image for alignment purposes.
- 3.2 Feature-based methods: This subsection explores methodologies that rely on identifying and matching specific features within the image.
-
4. Step2: Feature matching
-
- 4.1 Area-based methods
- 4.2 Feature-based methods
-
- 1. Introduction
-
5. Step three: Estimate the model structure
-
6. Step four: Conduct image resampling and transformation processing
1. Introduction
Image registration refers to the procedure of superimposing multiple images of the same scene captured at different times, angles, or sensor types. This operation involves geometrically aligning pairs of images, specifically a reference image and a sensed image. The discrepancies observed between these images arise primarily from variations in imaging conditions.
Image alignment serves as an integral part of numerous image analysis processes where final insights are derived through the integration of diverse data sources. Such as in applications like image fusion, change detection, and multichannel image restoration.
Usually registration is necessary for remote sensing applications such as multispectral classification environmental monitoring change detection image mosaicing weather forecasting creation of super-resolution images integration of data into geographic information systems GIS etc. In medical fields it involves combining CT scans and MRI data for comprehensive patient information monitoring tumor growth verifying treatments comparing patient data with anatomical references. Cartography requires map updating while computer vision focuses on target localization and automatic quality control among other tasks.
2. Image registration methodology
The applications of this method are typically categorized as four primary categories based on the method of image acquisition.
Multiple viewpoints (multiview analysis) are employed to capture images from various angles. Same-scene images are captured using multiple viewpoints. The objective is to achieve either a larger two-dimensional perspective or a three-dimensional reconstruction of the scanned scene.
不同时间点（多时相分析）
Various sensors, particularly known as multimodal analysis, are employed to capture images of the same scene through diverse sensing modalities. The objective of this process is aimed at integrating the information from various source streams to create a more comprehensive and detailed scene reconstruction.
Scene-to-model registration involves aligning images captured from both the actual scene and its digital representation. A computer-generated model, such as those used in GIS systems with DEMs or other similar datasets like another patient's imagery or an average specimen, serves as the reference for this alignment. The primary objective is twofold: first, to determine the precise location of the image within the registered model; second, to analyze similarities or differences between corresponding parts.
[four steps]
Feature detection. （特征检测）Salient and distinctive objects (closed-boundary regions, edges, contours, line intersections, corners, etc.) are manually or, preferably, auto-matically detected（突出和独特的物体被手动或自动检测）. For further processing, these features can be represented by their point representatives (centers of gravity, line endings, distinctive points), which are called control points (CPs)（控制点） in the literature.
该方法涉及特征匹配。（Feature Matching）该步骤通过确定感测图像与参考图像中检测到的特征之间的对应关系来实现。（In this step, ...）为了这一目的，采用了多种基于特征描述符、相似性度量以及它们之间空间关系的方法。（Various feature descriptors and similarity measures along with spatial relationships among the features are used for that purpose.）
转换模型估算。
形式化地确定映射函数的类型和参数，在将被感测图像与参考图像配准的过程中进行评估。
通过建立特征对应关系来确定映射函数的参数。
Image undergoes re-sampling and conversion.（图像经过重新采样和转换） The image undergoes transformation via mapping functions.（通过映射函数对图像进行变换） The image values at non-integer coordinates are calculated using suitable interpolation techniques.（通过适当的方法计算非整数坐标处的图像值）
3. Step1: Feature detection
3.1 Area-based methods
无feature detection
3.2 Feature-based methods
Region features.
Line features
Point features
4. Step2: Feature matching
4.1 Area-based methods
Corresponding correlation techniques
Fourier transform-based techniques
Information-theoretic measures
Optimization techniques
4.2 Feature-based methods
Various methodologies that utilize spatial relationships for analysis or modeling. Different techniques that rely on invariant descriptors to achieve robustness or uniqueness. Various relaxation-based techniques that employ iterative processes to reach equilibrium or optimal solutions. Multi-scale structures such as pyramids and wavelet transforms are fundamental in various applications.