Study On Image Fusion Algorithm Based On Neural Network And Guided Filter

Image fusion technology combines two or more different images into a higher quality image.The fused image has richer content and visual information,and is more suitable for human visual sense or computer processing.This technology has been widely used in medical,aerospace,remote sensing and other technical fields.The image fusion algorithm based on neural network and guided filter is discussed in this thesis.The specific research contents are as follows:For the transformation domain image fusion algorithm,the fixed transformation method is often used for the fusion image,and the fixed fusion rule is often used when the transformation coefficients are fused after the transformation.This thesis regards the deconvolution neural network as a transformation method.According to the content of the image to be fused,a transformation method is adaptively designed so that the image to be fused has its corresponding optimal transformation,which can best express the content of the image.According to the form of fusion rules to be adopted,Butterworth low-pass and high-pass filter banks are designed and used as the initial filters of the deconvolution neural network.To learn and train the deconvolution neural network in order to obtain the best transformation method.According to different images to be fused,an experimental study was conducted on the combination method of Butterworth low-pass and high-pass filter cutoff frequencies,and in the light of different evaluation parameters,the optimal combination of filter cutoff frequencies for the corresponding images was obtained.Using the designed deconvolution neural network,the optimal transformation of the image to be fused is performed to obtain the low-frequency and high-frequency feature maps.The feature maps are processed by the guided filter.Under the content of different images to be fused,the different fitness functions of genetic algorithms are designed,and the parameters of the guided filter are optimized to obtain the best fused image.Using the algorithm proposed in this thesis,the medical image fusion,multi-focus image fusion,and remote image fusion are studied experimentally.In the medical image fusion,15 groups of initial Butterworth filters are selected to learn and train the deconvolution neural network,15 groups of medical images are fused.The experimental results are compared with the guided filter image fusion algorithm based on wavelet transform.The fusion images obtained by this algorithm are all clearer and have better effect.In the multi-focus image fusion,15 groups of initial Butterworth filters are selected to learn and train the deconvolution neural network,12 groups of multi-focus images are fused.The experimental results are compared with the guided filter image fusion algorithm based on wavelet transform.Except two groups images have the ordinary experiments results,the rest fusion images obtained by this algorithm have higher quality.In the remote sensing image fusion,15 groups of initial Butterworth filters are selected to learn and train the deconvolution neural network,and 12 groups of remote sensing images are fused.The experimental results are compared with the guided filter image fusion algorithm based on wavelet transform.The fusion images obtained by this algorithm have higher quality except for three groups.