Comparison of Widely-used Image Restoration Methods for Additive and Multiplicative Noises

This paper evaluates and compares the performance of various image denoising techniques, including wavelet transform, Wiener filtering, non-local means (NLM), denoising autoencoders (DAEs), and feedforward autoencoders. There are cases of additive and multiplicative noises considered where the Gaussian distribution with zero mean stands for the first case and an exponential distribution with unity mean describes the second case (speckle). The experiments have been conducted on the given dataset which contained both still test images and synthetic aperture radar (SAR) images with varying noise levels to analyze the effectiveness of these denoising approaches. The performance of each method is assessed using key image quality metrics: Peak Signal-to-Noise Ratio (PSNR), Structural Similarity Index (SSIM), Mean Squared Error (MSE), and Normalized Absolute Error (NAK). The results demonstrate that denoising autoencoders DAEs provide superior performance in terms of structural preservation and noise reduction, particularly in low-noise conditions, while Wiener filtering proves to be computationally efficient and robust under high-noise conditions. These findings offer insights into the selection of effective denoising techniques for noisy images based on specific noise characteristics and application requirements.