研究目的
To enhance the performance of remote sensing image retrieval (RSIR) by introducing a new relevance feedback (RF) method called circular relevance feedback (CRF) that uses active learning algorithms to automate sample selection and integrate contributions through a circular fusion scheme.
研究成果
The proposed circular relevance feedback (CRF) method effectively enhances the performance of remote sensing image retrieval by automating sample selection through active learning algorithms and integrating their contributions in a circular fusion manner. Experimental results show improvements in average retrieval precision and recall compared to initial results and existing RF methods, demonstrating its utility and robustness. Future work could focus on optimizing the selection of AL algorithms and extending the method to broader datasets.
研究不足
The method relies on the quality of initial retrieval results from existing RSIR methods. The circular fusion scheme may increase computational complexity, and the selection of AL algorithms and their parameters could be optimized further. The dataset is limited to a specific LULC archive, which may not generalize to other types of remote sensing images.
1：Experimental Design and Method Selection:
The methodology involves using active learning (AL) algorithms to select samples from initial retrieval results automatically, replacing manual selection in relevance feedback (RF). Different AL algorithms are chosen to ensure representativeness and informativeness, and their contributions are integrated via a circular fusion scheme. A binary classifier (SVM) is used for reranking.
2：Sample Selection and Data Sources:
The dataset used is a land-use/land-cover (LULC) remote sensing image archive from the University of California at Merced, containing 2100 high-resolution aerial images classified into 21 categories. Images are 256x256 pixels with one-foot resolution.
3：List of Experimental Equipment and Materials:
No specific equipment or materials are mentioned; the focus is on computational methods and algorithms.
4：Experimental Procedures and Operational Workflow:
Initial retrieval results are obtained using existing RSIR methods (e.g., RSIR-BOW, RSIR-RFM, RSIR-DCNN). CRF is applied with multiple AL algorithms in a circular manner: output of one RF process serves as input for the next. SVM is trained with kernel functions chosen based on the similarity measure of the RSIR method. The process involves 10 iterations with 5 samples selected per iteration.
5：Data Analysis Methods:
Performance is evaluated using average retrieval precision (ARP) and average retrieval recall (ARR), calculated as nc/nt and ns/nt, respectively, where nt=20 (top 20 results). Statistical analysis includes comparisons with other RF methods.
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
