Author Affiliations
College of Science, Nanjing University of Aeronautics and Astronautics, Nanjing 211100, Jiangsu , Chinashow less
Fig. 1. Structure diagram of Yolov3 network
Fig. 2. Working mode of standard convolution
Fig. 3. Structure diagram of ghostbottleneck module
Fig. 4. Working mode of PW convolution
Fig. 5. Working mode of DW convolution
Fig. 6. Structure of Yolov3-new network
Fig. 7. Working mode of DSC
Fig. 8. Test system diagram of UAV charging experiment
Fig. 9. Training loss chart of Yolov3, Yolov3-new, and Yolov3-tiny
Fig. 10. UAV target training loss chart of Yolov3, Yolov3-new, and Yolov3-tiny
Fig. 11. Detection results of three algorithms on UAV target data set. (a) Yolov3; (b) Yolov3-new; (c) Yolov3-tiny
Performance | Class | Yolov3 | Yolov3-tiny | Yolov3-new |
---|
AP /% | Aeroplane | 82.52 | 80.37 | 77.75 | Bicycle | 93.64 | 89.58 | 91.57 | Bird | 87.67 | 74.93 | 83.80 | Boat | 89.70 | 71.26 | 90.02 | Bottle | 75.44 | 60.61 | 68.25 | Bus | 93.05 | 88.07 | 96.00 | Car | 87.25 | 84.90 | 88.29 | Cat | 93.27 | 87.23 | 89.51 | Chair | 84.63 | 71.48 | 76.48 | Cow | 92.38 | 80.05 | 91.96 | Diningtable | 88.56 | 78.01 | 79.53 | Dog | 92.18 | 85.36 | 81.50 | Horse | 94.06 | 86.50 | 93.57 | Motorbike | 92.09 | 79.73 | 85.87 | Person | 84.23 | 75.30 | 79.08 | Pottedplant | 79.44 | 56.36 | 73.15 | Sheep | 65.46 | 71.83 | 81.84 | Sofa | 97.39 | 91.56 | 89.77 | Train | 97.08 | 89.34 | 93.17 | Tvmonitor | 86.10 | 80.66 | 91.89 | mAP /% | | 87.81 | 75.16 | 85.15 |
|
Table 1. Performance comparison of three algorithms for voc2007 data set detection
Algorithm | Weight /MB | mAP /% | Speed /(frame·s-1) |
---|
Yolov3 | 236.0 | 87.81 | 17 | Yolov3-tiny | 33.8 | 75.16 | 47 | Yolov3-new | 29.7 | 85.15 | 33 |
|
Table 2. Comparison of detection performance of three algorithms
Algorithm | Weight /MB | AP /% | Speed /(frame·s-1) |
---|
Yolov3 | 236.0 | 90.17 | 17 | Yolov3-new | 29.7 | 89.33 | 33 | Yolov3-tiny | 33.8 | 0 | 47 |
|
Table 3. Performance comparison of three algorithms in project data set detection