研究目的
To explore the usefulness of image visibility graphs (IVG/IHVGs) in image processing and image classification, demonstrating their ability to encapsulate structural information and serve as effective filters and feature extraction methods.
研究成果
Image visibility graphs (IVG/IHVGs) provide a robust and efficient method for image filtering and feature extraction, achieving high classification accuracies in various tasks. Global features from IHVG and local features from IVG are particularly informative. The methodology is universal and computationally scalable, with potential for further enhancements using higher-order patches and multiplex network analysis.
研究不足
The study focuses on low-order visibility patches (p=3), which may not capture all complexities; higher orders could improve performance. The approach is tested on specific datasets and may not generalize to all image types. Computational efficiency, while linear, might be constrained by image size. Multiplex features for RGB images are not fully explored in terms of inter-layer correlations.
1：Experimental Design and Method Selection:
The study employs image visibility graphs (IVG and IHVG) to map images into graphs, using graph-theoretical analysis for feature extraction. Methods include defining IVG/IHVG, extracting global features like degree distribution, and introducing local features such as Visibility Patches.
2：Sample Selection and Data Sources:
Datasets include the Kylberg Texture Dataset (material textures), 2D HeLa (bio-medical textures), and natural texture datasets (Upstex, Brodatz Colored, Multiband). Images are grayscale or RGB, with specific sizes and classes.
3：List of Experimental Equipment and Materials:
Computational resources (e.g.,
4：5GHz IntelCore i7 processor), software (Matlab for classification), and standard image datasets. Experimental Procedures and Operational Workflow:
Steps involve mapping images to IVG/IHVG, computing graph properties (e.g., degree, clustering), applying filters, extracting features, and using classifiers (e.g., SVM, KNN) with cross-validation.
5：Data Analysis Methods:
Statistical analysis includes PCA for dimensionality reduction, calculation of classification accuracy, AUC, and model average accuracy. Algorithms for feature extraction (e.g., PatchProfile) are implemented with linear time complexity.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
