Multi-scale Extreme Exposure Images Fusion Based on Deep Learning

Existing multi-exposure fusion focus on fusing more than two images differently exposed. However, when there are only two images with extreme exposures (large exposure and low exposure), it is difficult to prevent relative brightness reversal from happening in the fused image. In this paper, we introduce a simple deep learning architecture for fusion of two images with extreme exposures. To obtain preferable features, the proposed algorithm considers both low and high resolution in the two input images with extreme exposures. Particularly, the two inputs are firstly decomposed to different Multi-scale layers using downsampling and convolutional neural network. The images are fused in different layers, and then the fused image is obtained by reconstructing using up-sampling and convolutional neural network. The quantitative and qualitative analysis of the experimental results show that the proposed algorithm outperforms existing multi-scale exposure fusion algorithms in the sense that it retains the natural brightness and improves the MEF-SSIM.