Fig. 1. The proposed network structure
Fig. 2. The operation process of channel shuffle
Fig. 3. Building blocks of ShuffleNetV2
Fig. 4. Spatial Attention Module
Fig. 5. SAM heatmap
Fig. 6. The structure of SSPP
Fig. 7. Feature map of SSPP
Fig. 8. PANet structure
Fig. 9. Improved PANet structure
Fig. 10. Sample images and labels in SSDD
Fig. 11. Mosaic data enhancement
Fig. 12. Target detection results of the improved algorithm in different scenarios
Fig. 13. P-R curve(the threshold changed from 0.05 to 0.95)
Fig. 14. Comparison of detection algorithms for complex background of near-shore ships
Fig. 15. Comparison of detection algorithms for small target ships in the open sea
Layer | Kernel size | Stride | Repeat | Output size | Output channels(1×) |
---|
Image | - | - | - | 224×224 | 3 | Conv1 MaxPool | 3×3 3×3 | 2 2 | 1 1 | 112×112 56×56 | 24 24 | Stage2 | - - | 2 1 | 1 3 | 28×28 28×28 | 116 | Stage3 | - - | 2 1 | 1 7 | 14×14 14×14 | 232 | Stage4 | - - | 2 1 | 1 3 | 7×7 7×7 | 464 | Conv5 | 1×1 | 1 | 1 | 7×7 | 1 024 | GlobalPool | 7×7 | - | - | 7×7 | - | FC | - | - | - | - | 1 000 | FLOPs | - | - | - | - | 146 M | Weights | - | - | - | - | 2.3 M |
|
Table 1. Overall architecture of ShuffleNetV2
SSPP | SAM | PAN | mAP/% | Speed/FPS |
---|
× | × | × | 85.6 | 52 | √ | × | × | 90.2 | 50 | √ | √ | × | 91.5 | 48 | √ | × | √ | 93.4 | 47 | √ | √ | √ | 94.7 | 46 |
|
Table 2. Accuracy comparison results of different modules
Algorithm | SSD | YOLOv3 | YOLOv4 | YOLOv3-tiny | YOLOv4-tiny | Ours |
---|
Near shore | Precision/% | 74.2 | 75.3 | 78.4 | 69.3 | 72.3 | 80.6 | Recall/% | 91.6 | 92.4 | 91.5 | 78.5 | 83.8 | 94.3 | Open sea | Precision/% | 87.3 | 92.3 | 96.7 | 87.5 | 91.2 | 97.8 | Recall/% | 93.4 | 95.4 | 98.2 | 83.4 | 85.6 | 97.2 |
|
Table 3. Comparison of different detection algorithms in near shore and open sea scenes
Algorithm | Backbone | FLOPs | Params | Speed/FPS | Model size/MB | mAP/% |
---|
SSD | VGG16 | 130.9 B | 99.6 M | 28 | 100 | 89.3 | YOLOv3 | Darknet53 | 156.3 B | 246.1 M | 26 | 236 | 92.3 | YOLOv4 | CSPDarknet53 | 128.5 B | 255.7 M | 27 | 244 | 95.6 | YOLOv3-tiny | Darknet53_tiny | 5.62 B | 8.86 M | 35 | 33.4 | 90.3 | YOLOv4-tiny | CSPdarknet53_tiny | 6.96 B | 6.06 M | 38 | 22.6 | 93.8 | Ours | ShuffleNetV2 | 1.52 B | 2.34 M | 46 | 5.3 | 94.7 |
|
Table 4. Performance comparison of different detection algorithms