Author Affiliations
Missile Engineering College, Rocket Force University of Engineering, Xi'an, Shaanxi 710025, Chinashow less
Fig. 1. Air-to-ground variable resolution scene. (a) High-altitude vision; (b) middle-altitude vision; (c) low-altitude vision
Fig. 2. Models. (a) SSD model; (b) FSSD model
Fig. 3. Scene-coupled multi-task object detection model
Fig. 4. Information activation module. (a) Synchronous activation; (b) asynchronous activation
Fig. 5. Scene-assisted multi-task datasets. (a) Scene- object datasets; (b) remote sensing scene datasets
Fig. 6. Visualization results of scene coupling multi-task model (VGG16) validation set
Fig. 7. Sequential scene change resolution target search. (a) Far view rasterization scene perception schematic; (b) high-altitude scene-aware guided object detection
Base model | Channel addition mAP /% | Channel concatenation mAP /% |
---|
SSD | 82.13 | 86.63 | FSSD | 86.78 | 90.45 |
|
Table 1. Two feature map channel fusion methods (synchronous activation, VGG16)
Base model | Object detection mAP /% | Scene classification mAP /% |
---|
SSD-IA | 86.63 | 98.21 | SSD-none | 83.12 | 98.31 | FSSD-IA | 90.45 | 98.69 | FSSD-none | 88.44 | 98.56 |
|
Table 2. Comparison of IA module on different framework models
Base model | Synchronous activation mAP /% | Asynchronous activation mAP /% |
---|
SSD | 86.63 | 84.31 | FSSD | 90.45 | 88.36 |
|
Table 3. Effect of synchronous and asynchronous activations on accuracy of object detection
Feature extraction | Object detection task | Scene classification task | Frame rate |
---|
AP /% | | precision /% |
---|
Car | | Truck | Airplane | Boat | Town | Airport | Waters | VGG16 | 91.97 | 77.51 | 98.42 | 93.91 | 98.32 | 98.75 | 99.01 | 30 | ResNet50 | 93.12 | 84.23 | 99.17 | 94.67 | 99.31 | 99.12 | 99.52 | 14 | MobileNetsv2 | 84.76 | 79.34 | 88.45 | 86.56 | 98.22 | 97.43 | 98.32 | 46 | Darknetv2 | 83.13 | 77.21 | 85.31 | 82.14 | 97.51 | 98.21 | 98.77 | 40 |
|
Table 4. Scene-coupled multi-task model detection results based on different feature extractions
Algorithm | AP | mAP |
---|
| Car | Truck | Airplane | Boat |
---|
SSD-VGG16 | 84.24 | 67.22 | 98.31 | 89.77 | 84.89 | FSSD-VGG16 | 88.87 | 69.45 | 97.62 | 92.28 | 87.05 | Proposed-VGG16 | 91.97 | 77.51 | 98.42 | 93.91 | 90.45 |
|
Table 5. Comparison of proposed algorithm with traditional object detection models %
Feature extraction | Precision /% |
---|
Town | Airport | Waters |
---|
VGG16 | 98.44 | 99.75 | 99.11 | ResNet50 | 98.51 | 98.98 | 99.43 | MobileNetsv2 | 97.32 | 97.93 | 98.92 | Darknetv2 | 97.73 | 97.66 | 98.57 |
|
Table 6. Classification results in remote sensing scenes under different feature extraction networks