Infrared and visible image fusion network based on low-light image enhancement and attention mechanism

Abstract The purpose of infrared and visible image fusion is to combine the information of different spectral imaging to improve the visual effect and information richness of the image. However, the visible images collected by the existing public datasets are often dim, and the fused images cannot fully depict the texture details and structure in the visible images. Moreover, most deep learning-based methods fail to consider the global information of input feature maps during the convolutional layer feature extraction process, which leads to additional information loss. To address these issues, this paper proposes an auto-encoder network that integrates low-light image enhancement with an adaptive global attention mechanism. First, a sharpening-smoothing balance model for low-light image enhancement is designed based on the Retinex model. Enhance the structure, texture, and contrast information of low-light images by adjusting the balance index of the model. Then, an adaptive global attention block is added to the auto-encoder network, which enhances features with important information by adaptively learning the weights of each channel in the input feature map, thereby improving the network's feature expression capabilities. Finally, in the fusion part of the auto-encoder network, a deep spatial attention fusion block is proposed to maintain the texture details in the visible image and highlight the thermal target information in the infrared image. Our experiments are validated on MSRS, LLVIP, and TNO datasets. Both qualitative and quantitative analyses demonstrated that our method achieved superior comprehensive performance compared to the state-of-the-art image fusion algorithms of recent years.