Fig 1.
The main framework of our proposed network.
Fig 2.
Structure diagram of VGG-16.
Fig 3.
The channel attention optimization module (CAOM).
Fig 4.
The feature fusion module(FFM).
Fig 5.
The feature aggregation module with channel attention(FAMCA).
Fig 6.
Train loss.
Fig 7.
Visual comparison of state-of-the-art satellite cloud image segmentation methods.
From top to bottom, (a) are 2 different satellite cloud images, (b) are the ground-truth corresponding to satellite cloud images, (c) are the results of Fmask [9], (d) are the results of SegCloud [17], (e) are the results of our method.
Table 1.
Comparison of results with state-of-the-art satellite cloud image segmentation methods.
Table 2.
Comparison of results of different backbone networks.
Table 3.
Ablation study for FFM module.
Table 4.
Ablation study for FCCRF module.
Fig 8.
Bad samples in segmentation results, (a) are the original image, (b) are the Ground-truth, (c) are the segmentation result of the proposed method.