Progressive fusion of hyperspectral and multispectral images based on joint bilateral filtering

Hyperspectral imaging (HSI) finds widespread applications across various fields, yet its utility is hindered by the low spatial resolution of the images. The fusion of low-resolution hyperspectral images (LR-HSI) with highresolution multispectral images (HR-MSI) offers the potential to obtain high-resolution hyperspectral images (HR-HSI). However, the current existing fusion strategies overlook the spatial-spectral loss caused by differences in image modalities, which prevents the full utilization of spectral and spatial information present in the source images. To tackle this challenge, a progressive network comprising two subnetworks has been proposed. The progressive structure facilitates feature extraction at multiple scales, enhancing the capability to extract features. The first part is the pre-training subnetwork with a focus on super spectral resolution prior. It aims to learn spectral information and preserve the spatial resolution of the multispectral images, effectively reducing modal differences and minimizing information loss. Furthermore, the fusion subnetwork integrates a joint bilateral filtering (JBF) module as a fusion strategy that leverages a traditional bilateral operator to effectively preserve edges, smooth noise, and enhance interpretability. Additionally, a supervised attention module (SAM) has been incorporated within the fusion subnetwork to establish the relationship between multispectral source images and fused images, thereby ensuring the preservation of spatial details in the fused image. Experimental results demonstrate that this method outperforms other state-of-the-art techniques in terms of both subjective visual evaluation and objective metrics.