Infrared image enhancement for transmission cable terminal equipment using detail-enhanced transformer CycleGAN

Abstract

This paper proposes a detail-enhanced transformer cycle generative adversarial network (DET-CycleGAN) to improve the quality of infrared images of transmission cable terminal equipment. In the proposed model, the CycleGAN network is constructed based on a U-Net structure and integrated with the transformer to achieve an overall improvement in image quality. In particular, a high-frequency detail enhancement module (DEM) that combines guided filtering and attention weighting is designed, which can effectively enhance the texture features of key areas. In addition, a novel synthetic dataset synthesis method is proposed by combining the atmospheric scattering model and histogram matching. This method uses the atmospheric scattering model to simulate the blur-induced degradation in real low-quality infrared images while using histogram matching for low-contrast feature adaptation, ultimately generating a physically plausible paired infrared image dataset. The model is first pre-trained on synthetic paired data and then fine-tuned using real data. Finally, through knowledge distillation, the model is made lightweight to facilitate deployment in resource-constrained environments. Experimental results show that compared to methods such as CycleGAN, the proposed method improves FID by 9.08%, BQS by 65.49%, saturation by 12.12%, colorfulness by 9.42% on average and reduces the parameters by about 83.5%.