关键词: 红外图像, CycleGAN网络, 卷积注意力模块, 跨模态转换

Abstract:

To address the problems in current infrared image generation such as insufficient contrast between target and scene, excessively large discrepancies from real scenes, indistinct thermal source features, and great difficulty in constructing measured infrared image datasets, an improved CycleGAN-based infrared image generation method was proposed. By optimizing the network structure of the generator and adding a non-local module and a CBAM convolutional attention mechanism into the generator, the extraction capability of CycleGAN for infrared features was enhanced, enabling the network to capture the subtle features of targets more accurately; a perceptual loss function was introduced to improve the detail retention and perceptual quality of the generated infrared images and enhance the stability of model training, realizing high-quality image conversion from visible light to infrared image. Experimental results demonstrate that the infrared images generated based on this method not only effectively improve image quality and highlight thermal source features but also significantly reduce the cost and difficulty of traditional infrared image acquisition, providing a more convenient and efficient solution for the acquisition and analysis of infrared images in practical applications.

Key words: infrared image, CycleGAN network, convolutional attention module, cross-modal conversion