Research on Real Scene Image Super-Resolution Based on Diffusion Model

Previous image super-resolution algorithms based on convolutional neural networks have primarily been trained on ideal synthetic datasets; however, their performance often declines significantly when applied to real-world scenarios. To effectively extract original information from real scene images, this paper proposes a generative diffusion model, Diffusion-Net, designed for real scene images, which integrates diffusion models into a U-shape network. Specifically, the backbone of this model consists of multiple layers of encoders and decoders, with processed images being fed into the diffusion model for further optimization. Additionally, we define a hybrid loss function to enhance the visual quality of the reconstructed images. To accelerate the model's training speed, a feature distillation module is designed. Furthermore, a multi-head attention mechanism is incorporated to enable the model to reconstruct high-quality texture details. The model is trained on the RealSR dataset, and experimental results demonstrate that, compared to state-of-the-art super-resolution models (such as RCAN and SwinIR), Diffusion-Net achieves superior performance across three key evaluation metrics—Peak Signal-to-Noise Ratio, Structural Similarity Index, and No-Reference Image Quality Assessment—while also exhibiting enhanced visual effects.