Figure 3.
Framework of the proposed IFANI for multiple image fusion tasks. Here we take multi-exposure image fusion, for example. The inputs contain N frames of images, which are processed by a Conv process to get N corresponding feature maps with c channels in each. The network has eight fusion blocks, each of which aggregates the most informative information among N feature maps by max-pooling and fuses this aggregated feature map with the N respective feature maps. The Conv module in each fusion block is composed of a convolutional layer, a ReLU, and a batch normalization layer, where the size of all convolutional kernels is .