首页|Region-aware discriminative learning GAN for super-resolution reconstruction of infrared imagery
Region-aware discriminative learning GAN for super-resolution reconstruction of infrared imagery
扫码查看
点击上方二维码区域，可以放大扫码查看
原文链接
NETL
NSTL
Elsevier
Infrared image super-resolution reconstruction is crucial for improving the quality of infrared images. The cutting-edge efforts often suffer from problems such as unnatural texture and fake details, undermining image realism. This study introduces a novel unified framework, which combines a multi-order degradation simulation model and region-aware discriminative learning with generative adversarial network, to train the optimal model for infrared super-resolution reconstruction. The multi-order degradation model is established to enhance the effectiveness of visual reconstruction by simulating different degradation instances in authentic infrared scenarios. The generator adopts residual-in-residual dense blocks to enhance detail preservation capability. The discriminator uses an encoder-decoder architecture for semantic discrimination. This enables simultaneous global and local assessments of images as real or fake, providing a more accurate discrimination of the generated image's authenticity. Additionally, to discriminate high-frequency artifacts and authentic details in infrared images with intricate textures, a region-aware discriminative learning strategy is introduced. Furthermore, a hybrid loss function is developed, integrating local loss, adversarial loss, pixel-wise and perceptual loss together for advanced adversarial training, making the reconstructed images more realistic and natural. The method's efficacy is demonstrated on simulated and real-world infrared image datasets, with comparative analysis showing significant improvements. Our model outperforms the state-of-the-art alternative solutions on multiple benchmarks.
NETL
NSTL
Elsevier
