Meta-filtration: Adaptive selection of multiple filter cascades for images with quality analysis

Abstract

This paper proposes a meta-filtration framework for denoising images corrupted by mixed noise (Gaussian, salt&pepper, speckle). Instead of fixed pipelines or AI-based methods, it uses a manually defined filter set and automatically selects combinations that improve image quality metrics (PSNR or MS-SSIM). The approach was tested on images of different sizes and mixed noise levels. Results show PSNR improvements of 7- 11dB, with MS-SSIM confirming preservation of fine details. An embedded implementation on a Zynq-7000 SoC achieved similar quality to MATLAB (within 0.1-0.8dB) and significantly reduced runtime, demonstrating practical efficiency on resourceconstrained hardware.