TY - GEN
T1 - Remote Sensing Image Fusion Method Based on Retinex Model and Hybrid Attention Mechanism
AU - Ye, Yongxu
AU - Wang, Tingting
AU - Fang, Faming
AU - Zhang, Guixu
N1 - Publisher Copyright:
© The Author(s), under exclusive license to Springer Nature Singapore Pte Ltd. 2024.
PY - 2024
Y1 - 2024
N2 - 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.
AB - 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.
KW - Channel attention mechanism
KW - Inverse Retinex model
KW - Pansharpening
KW - Remote sensing image fusion
KW - Spatial attention mechanism
UR - https://www.scopus.com/pages/publications/85190650108
U2 - 10.1007/978-981-97-1568-8_7
DO - 10.1007/978-981-97-1568-8_7
M3 - 会议稿件
AN - SCOPUS:85190650108
SN - 9789819715671
T3 - Communications in Computer and Information Science
SP - 68
EP - 82
BT - Space Information Networks - 7th International Conference, SINC 2023, Revised Selected Papers
A2 - Yu, Quan
PB - Springer Science and Business Media Deutschland GmbH
T2 - 7th International Conference on Space Information Network, SINC 2023
Y2 - 12 October 2023 through 13 October 2023
ER -