A novel fusion method of multispectral image and panchromatic image based on nonsubsampled contourlet transform（NSCT） and non-negative matrix factorization（NMF） is presented,the aim of which is to preserve both sp...A novel fusion method of multispectral image and panchromatic image based on nonsubsampled contourlet transform（NSCT） and non-negative matrix factorization（NMF） is presented,the aim of which is to preserve both spectral and spatial information simultaneously in fused image.NMF is a matrix factorization method,which can extract the local feature by choosing suitable dimension of the feature subspace.Firstly the multispectral image was represented in intensity hue saturation（IHS） system.Then the I component and panchromatic image were decomposed by NSCT.Next we used NMF to learn the feature of both multispectral and panchromatic images＇ low-frequency subbands,and the selection principle of the other coefficients was absolute maximum criterion.Finally the new coefficients were reconstructed to get the fused image.Experiments are carried out and the results are compared with some other methods,which show that the new method performs better in improving the spatial resolution and preserving the feature information than the other existing relative methods.展开更多
In the field of infrared and visible image fusion,researchers have put increasingly complex fusion networks forward to pursue better fusion metrics.This has led to a growing number of parameters in fusion models.Addit...In the field of infrared and visible image fusion,researchers have put increasingly complex fusion networks forward to pursue better fusion metrics.This has led to a growing number of parameters in fusion models.Additionally,most fusion models rarely address the issue of preserving background details in images,while these details are vital to subsequent advanced visual tasks,such as image analysis and recognition.In response to these limitations mentioned above,this paper proposes a novel image fusion algorithm called lightweight multi-scale hierarchical dense fusion network(LMHFusion).Concisely,we propose a lightweight multi-scale encoder.It extracts multi-scale features from input images through four encoding blocks with different receptive fields.Then,a designed hierarchical dense connection method is employed to concatenate distinct scale features.Unlike traditional manual fusion strategies,our fusion network is designed to be learnable and has adjustable weights.Moreover,we have specially designed a histogram equalization loss to train LMHFusion.This new loss produces fused images that contain both prominent structures and rich details.Through comparative analysis of LMHFusion and twelve other representative fusion models,it has been proven that LMHFusion can make the model more suitable for resource-constrained scenarios apart from enhancing the quality and visual effects of fused images.Our model is nearly 5000 times smaller in size compared to RFN-Nest.展开更多
基金Supported by the National Natural Science Foundation of China(60872065)
文摘A novel fusion method of multispectral image and panchromatic image based on nonsubsampled contourlet transform（NSCT） and non-negative matrix factorization（NMF） is presented,the aim of which is to preserve both spectral and spatial information simultaneously in fused image.NMF is a matrix factorization method,which can extract the local feature by choosing suitable dimension of the feature subspace.Firstly the multispectral image was represented in intensity hue saturation（IHS） system.Then the I component and panchromatic image were decomposed by NSCT.Next we used NMF to learn the feature of both multispectral and panchromatic images＇ low-frequency subbands,and the selection principle of the other coefficients was absolute maximum criterion.Finally the new coefficients were reconstructed to get the fused image.Experiments are carried out and the results are compared with some other methods,which show that the new method performs better in improving the spatial resolution and preserving the feature information than the other existing relative methods.
基金supported by the National Key Laboratory of Air-based Information Perception and Fusion and the Aeronautical Science Foundation of China(Grant No.20220001068001)。
文摘In the field of infrared and visible image fusion,researchers have put increasingly complex fusion networks forward to pursue better fusion metrics.This has led to a growing number of parameters in fusion models.Additionally,most fusion models rarely address the issue of preserving background details in images,while these details are vital to subsequent advanced visual tasks,such as image analysis and recognition.In response to these limitations mentioned above,this paper proposes a novel image fusion algorithm called lightweight multi-scale hierarchical dense fusion network(LMHFusion).Concisely,we propose a lightweight multi-scale encoder.It extracts multi-scale features from input images through four encoding blocks with different receptive fields.Then,a designed hierarchical dense connection method is employed to concatenate distinct scale features.Unlike traditional manual fusion strategies,our fusion network is designed to be learnable and has adjustable weights.Moreover,we have specially designed a histogram equalization loss to train LMHFusion.This new loss produces fused images that contain both prominent structures and rich details.Through comparative analysis of LMHFusion and twelve other representative fusion models,it has been proven that LMHFusion can make the model more suitable for resource-constrained scenarios apart from enhancing the quality and visual effects of fused images.Our model is nearly 5000 times smaller in size compared to RFN-Nest.
