Abstract: ‍ ‍At present， infrared imaging technology has been widely used in medicine， security， environmental monitoring， military detection， and other aspects. However， due to the inherent defects of low-cost infrared imaging equipment and the influence of the atmospheric environment on thermal radiation conduction， the acquired images have dark brightness， blurred details， and low contrast， which affects subsequent image semantic analysis and target detection and recognition. Traditional model-based infrared image enhancement methods often require image prior information， model parameters are related to the scene， and the model generalization ability is weak. The infrared image enhancement algorithm based on deep learning enhances the infrared image quality， but the structure is redundant， which is not conducive to edge deployment. Generative adversarial networks （GAN） can significantly enhance infrared image quality via rotational adversarial training of the discriminator and generator. However， this method entails substantial network training parameters， consumes considerable resources in edge deployment， and possesses high computational complexity. In this study， a lightweight infrared image enhancement network based on adversarial generation is designed， which improves the feature extraction efficiency and reduces the number of network layers by adding a multi-level feature fusion structure and designing a multi-scale loss function based on the GAN model， which improves the image quality and enhancement efficiency， which is conducive to the edge deployment of the algorithm. Experiments show that the proposed method considers the global and local features of the image by adding a multi-level feature fusion structure and a multi-scale loss function under the same number of parameters， ensuring that the detailed information is not lost and the computational complexity is not significantly increased under the premise of improving network performance. For similar infrared image enhancement effects， the number of model parameters is reduced by 75.0%， and the inference time of edge devices is reduced by 32.07%.