Infrared and Visible Image Fusion Using Detail Enhanced Channel Attention Network

Abstract

Fusion of infrared and visible images aims to maintain thermal radiation information and detailed texture information on a single image. Previous deep learning based methods require complex architecture of networks to extract features of both sources. In these methods, convolution filters act equally on each channel so the feature maps containing different brightness and gradient information are treated equally across channels, which reduces the representational ability of networks. In this paper, we innovatively introduce channel-attention mechanism to the fusion network. We propose a detail enhanced channel attention(DECA) block and apply it to the fusion network. DECA block takes into account the average of brightness and gradient information to rescale the feature maps. It allows network selectively emphasize useful features and suppress less useful ones. The proposed fusion network has relatively simple architecture and low operational complexity. This end-to-end network can generate fused images directly from source images. The experimental results on fusion of infrared and visible images demonstrate that our proposed method outperforms the state-of-the-art fusion methods.