Unsupervised Multi Focus Image Fusion based on Deep Image Prior

Multi-focus image fusion is an effective method to eliminate the defocusing blur in the imaging process. At present, due to the lack of real supervised data, most multi-focus image fusion algorithms can only synthesize datasets through artificial blur, which may lead to the loss of detailed features. According to Deep Image Prior, utilizing the implicit prior information of neural networks can solve many standard inverse problems. Motivated by this idea, this paper proposes a new unsupervised MFIF method, which uses two deep prior networks to model the fused image and the focus map respectively, then use the edge reconstruction loss function to constrain the training process. The edge reconstruction loss can compensate for the shortcomings in generating deep image prior details. In order to verify the feasibility of the algorithm, four widely used objective evaluation metrics were selected for performance evaluation in the experimental verification stage.