Optimized edge-aware frequency-guided filtering for robust image denoising

Abstract

The problem of denoising intrusion is still of great concern in computational imaging because of the trade-off between noise reduction and image structure and details recovery. This paper proposes an optimized edge-aware fast adaptive guided filter (E-FAGF) combining wavelet-domain decomposition, edge-awareness, and lightweight deep learning for efficient and effective denoising. The biorthogonal wavelet transform is employed to decompose noisy images into low- and high-frequency sub bands and an improved edge-attention map for selective high-frequency denoising. Regularization parameters are estimated pixel-wise by a compact convolutional neural network (CNN), allowing spatial-varying filtering to be done with multi-scale processing. The resultant E-FAGF consistently outperforms the state of the art on this dataset: on BSD500 for speckle and Gaussian noise (peak signal-to-noise ratio (PSNR) of 39.63 dB and 33.97 dB, respectively), and competitive performance for Poisson noise (30.84 dB) a large margin compared to the reference bilateral and non-local means. Our method maintains high structural similarity (up to 0.97 in structural similarity index measure (SSIM)), runs at 0.015 seconds per 512×512 image on graphics processing unit (GPU), and can be applied without dataset specific training. These results suggest the possibility of E-FAGF to achieve a balance between classical efficiency and learning-based adaptability, thereby forming a new scenario to combine fast and reliable image restoration for actual scenarios.