Research Of Multi-Focus Image Fusion Algorithm

Image fusion plays an important role in our current life,and it has important research meaning because it can fuse information that comes different images to a fused image.MultiFocus image fusion can take fuses a series of image that has different focus part.To the variety character of multi-focus image,three image fusion methods are proposed in the dissertation,and the main work is shown as follows:(1)To overcome the disadvantage between lack of detail protection and discontinuity of structure,a multi-image fusion method based on image cartoon && Texture decomposition and convolve sparse representation is proposed.Firstly,cartoon and texture image decomposition is applied to source images to obtain cartoon and texture parts.Then,cartoon parts fuse by using convolve sparse representation,dictionary learning is used to fuse texture parts.Finally,the cartoon part and texture part is fused.The experiment result shows that our method can overcome the disadvantage between lack of detail protection and discontinuity of structure,and has a better visual effect.(2)In the dissertation,a multi-image fusion method based on the texture part and PulseCouple Neural Network(PCNN)is proposed to solve that exists not accurate fusion edge in some fusion method.Firstly,cartoon and texture image decomposition is applied to source images to obtain texture parts.Then,texture parts are processed in PCNN,ignition maps can be obtained.Finally,after removed small noise pixels and smoothed by using Morphological Image Processing and bilateral filtering,the initial decision map becomes a smooth decision map,and fused images can be obtained with source images.The experiment result shows that our method can protect the original information,edge information,and clarity of source images,and it has a better performance.(3)This dissertation describes a MFIF-GAN(U-net)model for multi focus image fusion to solve the problem feature extraction and fusion rule.The generator is composed of U-net and attention mechanism model of spatial compression and stimulation.It mainly completes the feature extraction of multi focus image,and completes the fusion of multi focus image through convolution layer and excitation function.The discriminator is composed of neural network basic layers such as convolution layer and Relu layer.The loss function is divided into generator loss and discriminator loss.The generator loss mainly includes gradient loss in feature extraction stage,confrontation loss in generator and mapping loss in fusion stage.The discriminator loss mainly includes confrontation loss.Experimental results show that MFIF-GAN(U-net)can effectively fuse multi focus images and has excellent performance.