Infrared image enhancement algorithm based on improved generative adversarial networks

Currently, infrared imaging technology is extensively utilized across various domains, including medicine, security, environmental monitoring, and military detection. Nonetheless, the limitations associated with low-cost infrared imaging devices, coupled with the influence of atmospheric conditions on thermal radiation transmission, often result in images characterized by low brightness, blurred details, and diminished contrast. These deficiencies adversely affect subsequent processes such as image semantic analysis and target detection and recognition. Traditional model-based methods for enhancing infrared images typically rely on prior image information, with model parameters being contingent on specific scenes, which consequently hampers their generalization capabilities. While deep learning-based algorithms for infrared image enhancement can enhance image quality, they frequently exhibit redundant architectures that are not optimal for edge deployment. Generative Adversarial Networks (GANs) have shown promise in significantly enhancing infrared images by employing a training process that alternates between discriminator and generator networks. However, the substantial number of parameters involved in GANs leads to considerable resource consumption during edge deployment and results in high computational complexity. This paper introduces a lightweight infrared image enhancement network grounded in adversarial generation. By integrating a multi-level feature fusion structure and formulating a multi-scale loss function within the GAN framework, the proposed approach enhances feature extraction efficiency while minimizing the number of network layers. This improvement not only elevates image quality but also streamlines the enhancement process, rendering the algorithm suitable for edge deployment. Experimental findings indicate that, under equivalent parameter conditions, the incorporation of a multi-level feature fusion structure and a multi-scale loss function enhances both global and local features, ensuring the preservation of detail information. The proposed model enhances network performance without a significant increase in computational complexity, achieving a 75.0% reduction in parameter size and a 32.07% decrease in inference time on edge devices, all while maintaining comparable performance in infrared image enhancement.