Abstract
Aiming at the problem that the remote sensing image quality evaluation models with manually extracted features lack robustness and generality, this paper proposes a 3D CNN-based architecture and nuclear power plant for accurate remote sensing image quality assessment. The model incorporates two sub-networks. The DSVL-based sub-network is employed to extract multi-scale, multi-direction and high-level features by layer-wise training. Afterwards, the extracted feature maps are fused as flowed as input data of the second sub-network, which is designed with 3D CNN architecture and nuclear power plant for remote sensing image quality assessment. Experimental results on remote sensing image quality database from the GeoEye-1 and WorldView-2 satellites show that the proposed model can optimally discover the essential features of the image and effectively extract the high-frequency information of each level of image, and has better overall quality assessment performance than the other state-of-the-art methods.