RADFNet: An infrared and visible image fusion framework based on distributed network

Abstract

Introduction: The fusion of infrared and visible images can improve image quality and eliminate the impact of changes in the agricultural working environment on the information perception of intelligent agricultural systems.

Methods: In this paper, a distributed fusion architecture for infrared and visible image fusion is proposed, termed RADFNet, based on residual CNN (RDCNN), edge attention, and multiscale channel attention. The RDCNN-based network realizes image fusion through three channels. It employs a distributed fusion framework to make the most of the fusion output of the previous step. Two channels utilize residual modules with multiscale channel attention to extract the features from infrared and visible images, which are used for fusion in the other channel. Afterward, the extracted features and the fusion results from the previous step are fed to the fusion channel, which can reduce the loss in the target information from the infrared image and the texture information from the visible image. To improve the feature learning effect of the module and information quality in the fused image, we design two loss functions, namely, pixel strength with texture loss and structure similarity with texture loss.

Results and discussion: Extensive experimental results on public datasets demonstrate that our model has superior performance in improving the fusion quality and has achieved comparable results over the state-of-the-art image fusion algorithms in terms of visual effect and quantitative metrics.

Keywords: distributed fusion; edge attention; image enhancement; intelligent agriculture; multiscale channel attention.