An end-to-end multi-scale network based on autoencoder for infrared and visible image fusion

Infrared and visible image fusion aims to obtain a more informative fusion image by merging the infrared and visible images. However, the existing methods have some shortcomings, such as detail information loss, unclear boundaries, and not being end-to-end. In this paper, we propose an end-to-end network architecture for infrared and visible image fusion task. Our network contains three essential parts: encoders, residual fusion module, and decoder. First, we input infrared and visible images to two encoders to extract shallow features, respectively. Subsequently, the two sets of features are concatenated and fed to the residual fusion module to extract multi-scale features and fuse them adequately. Finally, the fused image is obtained by the decoder. We conduct objective and subjective experiments on two public datasets. The comparison results with the state-of-art methods prove that the fusion results of the proposed method have better objective metrics and contain more detail information and more explicit boundary.