研究目的
To develop a parallel implementation of attribute pro?les in CUDA for hyperspectral image classi?cation, aiming to reduce execution time compared to CPU implementations.
研究成果
The GPU implementation of extended attribute profiles significantly reduces execution time (up to 48x speedup) compared to CPU versions, enabling real-time processing of hyperspectral images. The method effectively models spatial information and improves classification accuracy, with potential for extension to other attributes and applications in remote sensing.
研究不足
The implementation is limited to specific attributes (area and standard deviation) and may not generalize to other attributes without modification. Memory constraints on GPU for large datasets require band-wise processing, which could affect efficiency. The study uses only two datasets, limiting generalizability.
1：Experimental Design and Method Selection:
The methodology involves a parallel implementation on GPU using CUDA, based on a flooding and filtering process without building the max-tree structure. It includes feature extraction using wavelets to reduce spectral bands and attribute filtering for spatial information modeling.
2：Sample Selection and Data Sources:
Two datasets are used: the Pavia Centre hyperspectral image (1096x715 pixels, 102 bands) and the Santiago area multispectral image (10748x12288 pixels, 8 bands).
3：List of Experimental Equipment and Materials:
A personal computer with an Intel Core i7 microprocessor, 16 GB RAM, and an Nvidia GTX 1070 GPU with 8 GB global memory. Software includes CUDA C++ and LIBSVM for classification.
4：Experimental Procedures and Operational Workflow:
Steps include feature extraction with wavelets, connected component labeling, attribute initialization and updating, merging regions, finding roots, and building attribute profiles. Kernels are executed on GPU with specific memory management (global, texture, constant, shared).
5：Data Analysis Methods:
Execution times are measured and compared between GPU and CPU implementations. Classification accuracy is assessed using SVM with RBF kernel and cross-validation.
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
