SR-BigGAN: lightweight image super-resolution with priors
摘要
Recent progress in the field of high-fidelity image synthesis using GANs has shown appealing outcomes, which motivates a series of successful image super-resolution (SR) works. However, most GAN-based SR models apply plain GAN models and are cumbersome compared to CNN-based SR models. In addition, the very advantage of conditional GAN has not yet been explored under the context of SR. In this paper, we develop a lightweight SR-BigGAN with priors for single-image super-resolution (SISR). First, our new model is an extension of BigGAN tailored to deep SR pipeline, retaining both generator and discriminator architectures, but with modifications to accommodate SR tasks. Second, prior knowledge defined as class labels from the low-resolution images, is fully leveraged through the conditional generative model to refine the SR process. Third, the lightweight nature of the model is achieved through knowledge distillation, focusing on reduced computational complexity and memory usage, making it the first practice of this kind in GAN-based SISR modeling. Extensive experiments on DIV2K, Pascal, mini-ImageNet, and SR benchmarks including Set5 and Set14 in an attempt to compare the Structural Similarity (SSIM), Peak Signal-to-Noise Ratio (PSNR) with the state-of-the-art models have shown appealing results. Our model achieves an average PSNR of 34.99 and SSIM of 0.791 across these datasets, demonstrating quantitative improvements over existing methods. The generated high-resolution images offer both perceptual enhancement and improved classification results. Additionally, explicit comparisons with GAN-based SR techniques such as ESRGAN and SRGAN highlight the superiority of our approach in both fidelity and efficiency. In particular, we achieve an average of