A Comparative Analysis of Spatial And Frequency Domain Smoothing Techniques For Digital Image Noise Reduction

In image acquisition and transmission, noise often hinders the integrity of digital images which are vital in different medical diagnostic applications, surveillance and remote sensing. This paper provides a comparative analysis of classical smoothing methods within the spatial and frequency domains with respect to digital image denoising. Spatial filters: Mean, Median, Gaussian, and Interquartile Range (IQR) as well as frequency-domain filters: IDEAL, Butterworth, and Gaussian Low-Pass filters were tested on prototypical test images corrupted with Gaussian (s=25) noise, salt-and-pepper (20% density) noise. The measures of performance used were: Peak Signal-to-Noise Ratio (PSNR), Structural similarity Index (SSIM), and Mean Squared Error (MSE), with the aid of qualitative visual examination. Findings show that there is a performance dichotomy. In salt-and-pepper noise, the nonlinear filters performed best in terms of PSNR (up to 32.1 dB) and SSIM (>0.85) were spatial, which removes impulsive noise by meeting edges. In the case of Gaussian noise, frequency-domain filters, most especially Butterworth Low-Pass filter (PSNR = 28.5 dB) offered smoothing that was superior, although with a loss of high-frequency detail, and (in the Ideal filter) with ringing artifacts. There were no single techniques that were observed to be optimal in all types of noise, and this highlights the significance of noise-sensitive filter choice. The work offers a validated parameter of selecting denoising algorithms in accordance with noise properties. The results are very compelling in the sense that adaptive or hybrid frameworks can be dynamically developed by integrating the strengths of the two domains. This analysis will be expanded in the future to color images, to which deep learning-based denoisers will be used as a reference, and context-aware filter selection intelligent systems will be explored.