Objective Multi-band composite imaging detection technology can fully leverage the advantages of detection in different bands. It not only acquires more target information but also has strong anti-interference ability. Intelligent image processing, as one of the key technologies of multi-band composite imaging, requires a large number of multi-band images for training. Therefore, how to obtain a large number of images of the same scene but in different bands quickly is a key issue that needs to be urgently solved at present. In order to efficiently and realistically convert visible light images into infrared images of different bands in the same scene, form a multi-band image dataset, and provide sufficient data support for the development and capability testing of multi-band composite detection algorithms, a multi-band infrared image generation method based on the improved StarGAN model is proposed in this paper.

Methods A study on the multi-band infrared image generation method is conducted in this paper based on the StarGAN model, which is capable of simultaneously learning multi-domain relationships and realizing mutual conversion. In order to improve the training efficiency of the model, the network structure of the StarGAN was improved to realize the simultaneous conversion of visible images to multi-band infrared images (Fig.2). The model training loss function was improved, and the difference in image features was adopted to replace the difference in pixel values. It can enhance the fidelity of the generated infrared images.

Results and Discussions The method proposed in this paper can convert visible light images into infrared images of different bands efficiently. The generated infrared images retain the scene feature information of the original visible images while having a relatively realistic infrared texture. For example, the generated short-wave infrared (SWIR) images can well reflect the strong SWIR radiation reflected by white clouds and vegetation, while the generated long-wave infrared (LWIR) images can well reflect the differences in infrared radiation emitted by various objects in the scene, showing the feature that the engines and tires of moving vehicles emit strong infrared radiation (Fig.6). Compare the output results of the method in this paper with those of Pix2Pix and CycleGAN methods on the same dataset. The proposed method is compared with three typical methods Pix2Pix, CycleGAN, and StarGAN under the same conditions. The experimental results show that the infrared images generated by the method proposed in this paper have a higher similarity to real infrared images and better image quality (Fig.7). In addition, the efficiency of model training has been significantly improved (Tab.1).

Conclusions A multi-band infrared image generation method based on the improved StarGAN model is proposed. The training efficiency of the improved StarGAN model is enhanced significantly. And the generated infrared image scene has clear structure and rich texture details. Comparing its performance with Pix2Pix, CycleGAN, and StarGAN, the following results can be obtained.The structural similarity index measure (SSIM) of SWIR images is improved by approximately 21%, 9% and, 10%. The learned perceptual image patch similarity (LPIPS) is improved by approximately 46%, 32% and 25%. For LWIR images, the SSIM is improved by approximately 19%, 13% and 8%. And the LPIPS is improved by approximately 56%, 49% and 37%. The proposed method has a significant improvement in model training efficiency and image generation quality compared with traditional methods. It has good application value.