Performance Comparison of Autoencoders and Traditional Denoising Methods Across Four Image Datasets: Metric-Based Optimization

Noise in images presents a significant challenge in applications requiring high quality images what leads to choosing and applying best restoration and compression methods. This paper considers comparison of the ability of several methods such as wavelet transform, denoised autoencoder, Wiener, feedforward autoencoder, and NLM (Non-Local Means) to compress noisy images. To analyze the performance of each method, evaluation is based on the typical metrics as PSNR (Peak Signal-to-Noise Ratio), SSIM (Structural Similarity Index), MSE (Mean Squared Error) and NAK (Normalized Absolute Error), to ensure a comprehensive analysis of the methods across various image types, including structurally uniform regions (homogeneous images), complex textures (DTD and texture images), and SAR (Synthetic Aperture Radar) images with speckle. The results have shown that autoencoders give better results because they tend to preserve structural details. On the other hand, Wiener filtering is computationally efficient and can be chosen for resource-constrained applications. These findings offer valuable guidance for selecting the most appropriate method based on both image characteristics and application-specific requirements and underscores the importance of selecting denoising methods tailored to the dataset and noise characteristics.