Abstract:To address the challenges of excessive smoothing and visual quality degradation in hot-rolled steel image super-resolution, we propose RL super-resolution generative adversarial networks (RL-SRGAN), an enhanced network incorporating several technical advances. The method features a receptive field-aware spatial pyramid pooling fast (R-SPPF) module that significantly improves global feature extraction for strip steel images, along with a relativistic average discriminator that enhances training stability. Additionally, we introduce a novel local contrast loss function to better preserve fine texture details during reconstruction. Extensive experiments demonstrate consistent improvements of 0.34 dB in peak signal-to-noise ratio (PSNR) and 0.014 in structural similarity (SSIM) over SRGAN on standard benchmarks (Set5, Set14, BSD100, and Urban100), while producing more visually realistic outputs. The method’s practical value is further validated on the NEU-DET industrial dataset, achieving 76.4% mean average precision (mAP) with 5.3% faster convergence, confirming its strong generalization capability for real-world inspection tasks.