KEDM: Knowledge-Embedded Diffusion Model for Infrared Image Destriping

Abstract

Infrared imaging systems are widely used across industries. However, their output images often exhibit striped noise due to the nonuniform response of the detection system, which significantly affects image quality and visual fidelity. To address challenges such as incomplete stripe removal, potential loss of image details and textures, and the generation of artificial artifacts during destriping, we propose a novel stripe removal method based on a knowledge-embedded diffusion model (KEDM). This approach effectively integrates the spatial distribution characteristics of stripe noise with an innovative, data-driven diffusion network model, creating a hybrid knowledge and data-driven framework for stripe correction. The core components of KEDM are the latent diffusion model (LDM) architecture and the directional wavelet convolution module (DWCM). Specifically, LDM leverages a pretrained variational autoencoder (VAE) to transform the input image into latent feature space for efficient diffusion propagation, reducing computational complexity while preserving image restoration quality. Meanwhile, DWCM uses wavelet convolution operations to construct prior loss functions for stripe noise, precisely guiding the diffusion reconstruction process to achieve a clean, stripe-free image. Empirical evaluations on several benchmark datasets demonstrate that the proposed KEDM outperforms other state-of-the-art destriping algorithms in terms of visual quality and quantitative metrics, validating its excellent performance.