As one of the key tools of damage assessment technology, infrared image is highly valued for its ability to provide reliable target information in various complex environments. However, due to the limited performance of the hardware equipment and the impact of the shooting environment, the infrared camera can not collect enough infrared image data to support the model training, which limits its effectiveness in practical applications. In contrast, the method of converting from visible to infrared images is becoming an alternative due to its low cost and ease of operation. Under the background of insufficient infrared data, this conversion method can effectively supplement the data source and support the damage assessment of military firing range. At present, the research on the deficiency of infrared data mainly focuses on the conversion from visible image to corresponding image. How to learn the mapping correlation between cross-modal data in the training process has become a key research problem that needs to be solved.In recent years, researchers have proposed a variety of cross-modal methods for converting visible light images to infrared images. By embedding the input visible image into the potential feature space, the corresponding infrared image is generated using the nonlinear transformation relationship. For example, the use of generative adversarial networks, unsupervised learning and self-supervised learning. However, the significant semantic gap between visible and infrared images remains a challenge. In order to effectively achieve cross-modal conversion, these methods often require additional constraints to ensure that the generated infrared image is not only close to realistic in appearance, but also able to retain key thermal characteristics. However, these methods show some limitations when dealing with the conversion task from visible light to infrared, because these methods are not designed for infrared images, they are difficult to achieve the expected effect when generating images that meet the infrared thermal characteristics. It is worth noting that the data collected in real scenarios is often unaligned, which increases the difficulty of cross-modal transitions.In order to solve the above problems, this paper proposes an infrared image generation method, Combining Generative Adversarial Network and Contrastive Learning for Infrared Image Generation (CLGAN), which combines generative adversarial network and contrast learning. Specifically, the method combines bidirectional contrast learning and feature mapping constraints to improve the quality of unpaired cross-mode conversion, effectively constrains cross-mode feature mapping, and ensures that the generated infrared image accurately reflects the thermal feature while retaining the visual feature. At the same time, the cross fusion self-attention is proposed to integrate local and remote associations adaptively with self-attention to realize efficient fusion of global information and ensure full utilization and comprehensive expression of feature information. In addition, the multi-scale feature refinement module is designed to further enhance the expression ability of key details and significantly improve the quality of the generated images through the refinement processing of multi-scale features. Extensive experimental results show that the proposed CLGAN outperforms existing methods in terms of both visual effects and performance indicators, generates clearer and more realistic infrared images, and has important potential in downstream tasks with limited infrared data.