Multi-exposure Image Fusion via Adaptive Multi-Branch Input.

In this paper, we propose an adaptive multi-exposure image fusion (AMEF) framework using self-supervised learning. Our encoder-decoder network can adaptively increase the number of feature extraction branches to fuse low dynamic range (LDR) images of arbitrary size and number without adding parameters. We incorporate channel and pixel attention mechanisms into the feature extraction branches to provide greater flexibility in processing image information with different exposures. By focusing more attention on complementary features, our framework can achieve better image fusion. In addition, we introduce a local residual channel attention module (LRCA) in the decoder, which assigns different weights to the fused features, extending the representation capability of convolutional neural networks (CNNs). Finally, We evaluate our method on a multi-exposure image fusion benchmark dataset (SICE) and compare it with recent deep learning-based methods objectively and subjectively. Three of our quantitative measures achieve the best results, and the remaining measures are close to existing algorithms.