Two-Stage Image Denoising with FFDNet and Non-Local Enhancement under Perceptual Evaluation

This paper proposes a two-stage hybrid image denoising framework that combines deep learning and traditional filtering techniques to improve visual quality under high Gaussian noise. In the first stage, an FFDNet-based convolutional neural network removes most of the noise with spatially adaptive strength, benefiting from its lightweight design and noise-level-aware input. In the second stage, we refine the denoised output using a modified Non-Local Means (NLM) filter to enhance texture details and suppress residual artifacts. To address the lack of reference images in real-world test sets, we employ the no-reference PIQE metric for perceptual quality evaluation. Experimental results demonstrate that our method achieves a balance between effective denoising and texture preservation, particularly under variable and strong noise levels.