LFAD: Locally- and Feature-Adaptive Diffusion based Image Denoising

LFAD is a novel locally- and feature-adaptive diffusion based method for removing additive white Gaussian (AWG) noise in images. The method approaches each image region individually and uses a different number of diffusion iterations per region to attain the best objective quality according to the PSNR metric. Unlike block-transform based methods, which perform with a predetermined block size, and clustering-based denoising methods, which use a fixed number of classes, our method searches for an optimum patch size through an iterative diffusion process. It is initialized with a small patch size and proceeds with aggregated (i.e., merged) patches until the best PSNR value is attained. Then the diffusion model is modified; instead of the gradient value, we use the inverse difference moment (IDM), which is a robust feature in determining the amount of local intensity variation in the presence of noise. Experiments with benchmark images and various noise levels show that the designed LFAD outperforms advanced diffusion-based denoising methods, and it is competitive with state-of-the-art block-transformed techniques; block and ring artifacts inherent to transform-based methods are reduced while PSNR levels are comparable.