Fig. 1. Buildings are characterized by visible light remote sensing in night light remote sensing images(a) Local structural similarity(b) Remote context
Fig. 2. Flow chart of detection network
Fig. 3. Four kinds of Classification network structure W
Fig. 4. Prior box matching correction
Fig. 5. 4 kinds of network detectors
Fig. 6. Sampling point graph of deformable convolution with unit number 9 passing through 3 layers
Fig. 7. Structure comparison between GC module and CGC module
Fig. 8. Two CGC connection modes
Fig. 9. RGB remote sensing picture and night light remote sensing picture(a) RGB of 123 object, (b)RGB of 4 object, (c) Sensing of 123 object, (d) Sensing of 4 object
Fig. 10. Sample picture of luminous remote sensing data set(object and category are marked in the picture)
Fig. 11. Network detection image(a)No module added(b)Add CGC2 module
Fig. 12. Different networks’ P-R curve
Fig. 13. Different networks’ F-measure
输入 | A网络 | B网络 | D网络 |
---|
| | | |
|
Table 1. Hidden layer characteristics of different networks
Table 2. Attention graphs of different global semantic modules
类别(单位) | 1 | 2 | 3 | 4 | 5 |
---|
数量/个 | 5283 | 1138 | 278 | 227 | 166 | 平均边长/像素 | 72 | 160 | 1010 | 318 | 269 | 平均PSNR/dB | 19.26 | 19.89 | 21.17 | 21.71 | 23.24 |
|
Table 3. All categories in the data set
分类网络 | 参数量(107) | mAP |
---|
A B C D | 2.563 750 4 2.522 928 0 2.191 003 2 0.636 686 4 | 0.387 9 0.389 6 0.419 9 0.404 0 |
|
Table 4. The performance of 4 kinds of networks on the night light remote sensing data set
阶段数 | 5 | 6 | 7 | 8 | 1-8 |
---|
mAP | 0.400 9 | 0.401 3 | 0.400 8 | 0.401 2 | 0.400 8 |
|
Table 5. Adding expansion convolution in different stages
阶段数 | 4 | 5 | 6 | 7 | 1-7 |
---|
mAP | 0.408 9 | 0.406 3 | 0.409 1 | 0.413 7 | 0.416 7 |
|
Table 6. Add GC module in different stages
实验 | CN | ECN | DCN | SSD匹配 | GE 匹配 | GC | CGC1 | CGC2 | mAP |
---|
1 | √ | | | √ | | | | | 0.383 3 | 2 | √ | | | | √ | | | | 0.387 6 | 3 | | | √ | √ | | | | | 0.398 2 | 4 | | | √ | | √ | | | | 0.404 0 | 5 | | √ | | | √ | | | | 0.400 8 | 6 | | | √ | | √ | √ | | | 0.416 7 | 7 | | | √ | | √ | | √ | | 0.424 5 | 8 | | | √ | | √ | | | √ | 0.429 6 |
|
Table 7. Comparison of ablation experiments of D-network modules
网络 | 1 | 2 | 3 | 4 | 5 | mAP | 帧数 |
---|
Yolov3 | 0.264 | 0.344 | 0.398 | 0.349 | 0.435 | 0.3580 | 6.7 | Fcos | 0.278 | 0.351 | 0.423 | 0.395 | 0.480 | 0.3854 | 5.6 | Faster R-CNN | 0.287 | 0.391 | 0.601 | 0.281 | 0.414 | 0.3949 | 4.3 | D | 0.302 | 0.350 | 0.457 | 0.401 | 0.510 | 0.4040 | 16.7 | CGC-D | 0.397 | 0.363 | 0.467 | 0.407 | 0.514 | 0.4296 | 14.3 |
|
Table 8. Different network detection effects