Remote Sensing Image Fusion Method Based on Retinex Model and Hybrid Attention Mechanism

Pansharpening is a technique that fuses a low-resolution multispectral image (LRMS) and a panchromatic image (PAN) to obtain a high-resolution multispectral image (HRMS). Based on the observation that PAN and LRMS respectively have the characteristics of illumination component and reflection component of HRMS after Retinex decomposition, this paper proposes an inverse Retinex model guided pansharpening network, termed as AIRNet. Specifically, a Spatial Attention based Illuminance Module (SAIM) is proposed to convert the PAN to the illuminance component of HRMS. And a Hybrid Attention-based Reflectance Module (HARM) is used to convert the LRMS to the reflection component of the HRMS. Finally, based on the inverse Retinex model, the corresponding illuminance component and reflection component of the obtained HRMS are fused to obtain HRMS. Qualitative and quantitative comparison experiments with state-of-the-art pansharpening methods on multiple remote sensing image datasets show that AIRNet has significantly outstanding performance. In addition, multiple ablation experiments also show that the proposed SAIM and HARM are effective modules of AIRNet for pansharpening.