Abstract:

In the multi-focus image fusion problem,in order to preserve as much original image information as possible and improve the quality of image fusion,a two-stage image fusion network based on unsupervised learning is designed by combining the variational autoencoder structure and the gray variance product function in the no-reference image clarity evaluation index.In the training phase,the multi-scale structural similarity is proposed as the loss function and the total deviation loss is introduced to suppress the noise in the image.An encoder-decoder network based on the variational autoencoder structure is constructed to train the original image reconstruction task.In the fusion stage,after using the trained encoder to encode the features of the fused image,the improved gray variance product function method is used to distinguish the clear pixels.The final decision map is generated by mathematical morphology optimization.Finally,the weighted fusion strategy is used to complete the final fusion of the image.Experimental results show that although this method uses fewer model parameters,it retains more original image information in the encoding and decoding process,and is superior to the traditional spatial frequency-based discrimination method in the pixel discrimination process.Compared with a variety of representative image fusion methods,the proposed method has achieved a superior fusion performance in both subjective and objective evaluation.

Key words: multi-focus image fusion, unsupervised learning, variational autoencoder, product of gray variance