研究目的
To propose a hierarchical image fusion framework that applies layer-by-layer deep learning techniques to explore the detailed information of images and extract key information of images for dictionary learning, aiming to overcome the difficulties of obtaining a complete and non-redundant dictionary in sparse representation-based image fusion.
研究成果
The proposed image fusion approach effectively explores the structure and detailed information of source images by analyzing their geometric structure and clustering them into smooth, stochastic, and dominant orientation patches. The constructed dictionary is both compact and informative. Comparison experiments demonstrate the proposed solution's superior performance in visible-infrared image fusion over other methods. Future work includes integrating de-noise functions and optimizing clustering and dictionary learning processes.
研究不足
The proposed solution, while effective, has room for improvement in integrating de-noise functions into the dictionary learning process to enhance noise robustness. The efficiency of clustering and dictionary learning, which decides the performance of the proposed solution, also requires further optimization.
1：Experimental Design and Method Selection:
The paper proposes a hierarchical image fusion framework integrating deep learning techniques for detailed and key image information exploration and extraction. It clusters source image patches into smooth, stochastic, and dominant orientation patches based on geometric similarities. High-frequency and low-frequency components of these patches are fused using max-L1 and L2-norm based weighted average fusion rules, respectively.
2：Sample Selection and Data Sources:
Twenty-five pairs of 256 × 256 visible-infrared images obtained from http://www.imagefusion.org are used for testing.
3：List of Experimental Equipment and Materials:
The experiments are implemented using MATLAB 2014a and Visual Studio 2013 Community on an Intel(R) Core(TM) i7-4720HQ CPU@ 2.60 GHz Laptop with 12.00 GB RAM.
4：60 GHz Laptop with 00 GB RAM.
Experimental Procedures and Operational Workflow:
4. Experimental Procedures and Operational Workflow: The framework consists of geometric classification of image patches, decomposition into high-low frequency components, and sub-dictionary learning and merging. The fused high-frequency and low-frequency components are combined to form the final fused image.
5：Data Analysis Methods:
Six mainstream objective evaluation metrics are implemented for quantitative evaluation, including edge intensity (EI), mutual information (MI), edge retention (QAB/F), visual information fidelity (VIF), Yang et al. proposed fusion metric (QY), and Chen-Blum metric (QCB).
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
