Author Affiliations
1Graduate Team, Engineering University of PAP, Xi'an , Shaanxi 710086, China2School of Information Engineering, Engineering University of PAP, Xi'an , Shaanxi 710086, China3School of Civil Engineering, Xinjiang University, Urumqi , Xinjiang 830000, Chinashow less
Fig. 1. Partial pictures of datasets. (a) Dataset A; (b) Dataset B
Fig. 2. Analysis of size of drone targets in two datasets. (a) Dataset A; (b) Dataset B
Fig. 3. Structure diagram of YOLOv4-tiny algorithm. (a) YOLOv4-tiny; (b) CSPBlock
Fig. 4. Structure diagram of DTD-YOLOv4-tiny model
Fig. 5. ShuffleNetV2 and improved backbone network structure. (a) ShuffleV2Block; (b) backbone network of ShuffleNetV2; (c) backbone network of proposed algorithm
Fig. 6. FPN structure comparison of different detection models. (a) YOLOv4-tiny; (b) YOLOv4-tiny (YOLO-Head enhancement); (c) DTD-YOLOv4-tiny
Fig. 7. Working principle of reorg_layer
Fig. 8. Working principles of sub-pixel Conv and sub-pixel. (a) Sub-pixel Conv; (b) sub-pixel
Fig. 9. Comparison of accuracy and detection speed of different target detection models under different datasets. (a) Dataset A; (b) Dataset B
Fig. 10. Comparison of partial detection results of test set on different datasets. (a) YOLOv4-tiny (Dataset A); (b) DTD-YOLOv4-tiny (Dataset A); (c) YOLOv4-tiny (Dataset B); (d) DTD-YOLOv4-tiny (Dataset B)
Algorithm | Image size | mAP@50 /% | GFLOPs | Detection speed /(frame·s-1) |
---|
YOLOv4 | 416 | 65.7 | 60.1 | 55 | YOLOv3-tiny | 416 | 33.1 | 5.6 | 345 | YOLOv4-tiny | 416 | 40.2 | 6.9 | 330 |
|
Table 1. Performance comparison of different algorithms in MS COCO dataset
Model | Head improve | | Neck improve | | Backbone Improve | AP /% | Parameters | GFLOPs |
---|
Anchor box imporve | YOLO-Head enhancement | ShuffleV2 Block(B) | | reorg layer | sub pixel | Shufflev2 Block(B) | | ShuffleNetv2 Backbone | Proposed Backbone |
---|
YOLOv4-tiny (Baseline) | | | | | | | | | | | 43.1 | 5.8×106 | 12.1 | Imporve A | √ | | | | | | | | | | 84.1 | 5.8×106 | 12.1 | Imporve B | √ | √ | | | | | | | | | 91.3 | 6.1×106 | 14.3 | Imporve C | √ | √ | | | √ | | | | | | 93.6 | 6.8×106 | 17.1 | Imporve D | √ | √ | | | √ | √ | | | | | 94.5 | 6.5×106 | 16.1 | Imporve E | √ | √ | | | √ | √ | | | √ | | 87.1 | 0.65×106 | 1.9 | Imporve F | √ | √ | | | √ | √ | | | | √ | 87.4 | 0.64×106 | 2.2 | Imporve G | √ | √ | √ | | √ | √ | | | | √ | 85.3 | 0.27×106 | 0.9 | DTD- YOLOv4-tiny | √ | √ | √ | | √ | √ | √ | | | √ | 89.4 | 0.32×106 | 1.1 |
|
Table 2. Ablation experiment results of DTD-YOLOv4-tiny algorithm
Algorithm | Image size | P /% | R /% | AP /% | Parameters | GFLOPs | GPU /GB |
---|
YOLOv3-SPP | 416×234 | 93.5 | 96.2 | 96.8 | 625×106 | 116.9 | 3.85 | YOLOv4 | 416×234 | 85.6 | 96.6 | 95.5 | 639×106 | 105.9 | 5.6 | YOLOv4-tiny | 416×234 | 81.5 | 87.6 | 84.1 | 5.87×106 | 12.1 | 0.7 | DTD-YOLOv4-tiny | 416×234 | 83.5 | 86.7 | 89.4 | 0.327×106 | 1.1 | 0.52 | DTD-YOLOv4-tiny | 608×342 | 88.3 | 93.9 | 94.2 | 0.327×106 | 1.1 | 0.583 | DTD-YOLOv4-tiny | 960×540 | 87.8 | 95.6 | 95.0 | 0.327×106 | 1.1 | 1.3 |
|
Table 3. 4 Performance comparison of different algorithms on Dataset A
Model | Image size | AP /% | Model size |
---|
YOLOv3[18] | 1333×800 | 84.9 | 234.1 MB | Fast RCNN(MobileNet)[18] | 1333×800 | 67.5 | 162.5 MB | Casacade RCNN(MobileNet)[18] | 1333×800 | 78.0 | 384.9 MB | TIB-Net[18] | 1333×800 | 89.2 | 697.0 KB | YOLOv4-tiny | 1344×756 | 78.5 | 23 MB | DTD-YOLOv4-tiny | 960×540 | 80.3 | 1.4 MB | DTD-YOLOv4-tiny | 1344×756 | 83.3 | 1.4 MB | DTD-YOLOv4-tiny | 1920×1080 | 85.1 | 1.4 MB |
|
Table 4. Performance comparison of different algorithms on Dataset B