Natural statistics of multisensor images: Comparative analysis and application to image classification and image fusion

This paper presents a comparative analysis of multisensor images, including long-wave infrared (LWIR), near-infrared (NIR), and visible light (VIS) images, using natural scene statistics (NSS) models with a focus on image classification and optimizing multisensor image fusion. The study considers the impact of several factors, such as flat region, light condition and noise power, on four coefficients, namely, mean subtracted contrast normalized (MSCN), paired product, log-derivative, and steerable pyramid. It reveals remarkable patterns in the NSS of LWIR, NIR, and VIS images, showcasing both similarities and differences. Notably, NIR and VIS images exhibit a high statistical similarity, while LWIR displays a decorrelated pattern compared to them. Additionally, two practical tasks are addressed: image classification and multisensor image fusion. The obtained results demonstrate that the NSS of LWIR, NIR, and VIS images can firstly be used to build image distortion classifiers for multisensor images. Secondly, they can be used in multisensor image fusion to select the optimal combination of images that maximize the texture information in the fused image. These findings offer valuable insights for improving sensor selection and image fusion quality assessment.