NLIFD: improved non-local mean filtering algorithm for infrared face image denoising

Abstract To address the challenge that infrared face images contain varying degrees of noise due to imaging mechanisms and external environmental factors, which adversely affect subsequent processing and analysis. we propose an improved non-local mean filtering-based denoising algorithm with three key innovations: (1) Development of a hybrid filtering equation combining diffusion and shock terms for preprocessing, effectively reducing strong noise interference in weight coefficient calculation; (2) Construction of a weight calculation framework through combined gradient features with structural similarity measurement using local information blocks, specifically designed for infrared facial characteristics; (3) Implementation of computational optimization via Euclidean distance symmetry analysis combined with adaptive filtering parameters, achieving enhanced processing efficiency. Additionally, we propose a novel dual-domain metric, termed Noise Binary Difference (NBD), which simultaneously evaluates noise suppression and structural fidelity preservation, thereby addressing the limitations of conventional single-dimensional indices. Experimental results demonstrate that the proposed improved algorithm outperforms the original non-local mean filtering algorithm and other denoising algorithms across various evaluation indices. Moreover, it significantly reduces the computational time for noise removal in infrared face images. The proposed evaluation index, Noises Binary Difference, provides a comprehensive and effective assessment of denoising algorithm performance and has considerable potential for broader applications.