Fig. 1. Yolov3 network structure
Fig. 2. Composite backbone network structure
Fig. 3. Feature augment block
Fig. 4. Yolo-C network structure of one-stage dual-network object detection algorithm
Fig. 5. Random X-ray images of GDXray dataset
Fig. 6. Random X-ray images of SIXray dataset
Fig. 7. Sample display of prohibited items
Fig. 8. Comparison graph of network training process
Fig. 9. IoU definition and calculation diagram
Fig. 10. Detection results of experiments 3, 5 and 7
Type | Layer | Filter number | Size | Output |
---|
DBL | Conv | 32 | 3×3 | 416×416 | Res1 | Conv2 | 64 | 3×3, 1×1 | 208×208 | — | Upsample | 32 | 3×3 | 416×416 | Res1' | Conv2 | 64 | 3×3, 1×1 | 208×208 | Res2 | Conv4 | 128 | 3×3, 1×1 | 104×104 | — | Upsample | 64 | 3×3 | 208×208 | Res2' | Conv4 | 128 | 3×3, 1×1 | 104×104 | Res8 | Conv16 | 256 | 3×3, 1×1 | 52×52 | — | Upsample | 128 | 3×3 | 104×104 | Res8' | Conv16 | 256 | 3×3, 1×1 | 52×52 | Res8 | Conv16 | 512 | 3×3, 1×1 | 26×26 | — | Upsample | 256 | 3×3 | 52×52 | Res8' | Conv16 | 512 | 3×3, 1×1 | 26×26 | Res4 | Conv8 | 1024 | 3×3, 1×1 | 13×13 | — | Upsample | 512 | 3×3 | 26×26 | Res4' | Conv8 | 1024 | 3×3, 1×1 | 13×13 | DBL | Conv | 30 | 3×3 | 13×13 | Head | — | — | — | — | FAB | Upsample | | 3×3, 1×1 | 26×26 | DBL | Conv | 30 | 3×3 | 26×26 | Head | — | — | — | — | FAB | Upsample | | 3×3, 1×1 | 52×52 | DBL | Conv | 30 | 3×3 | 52×52 | Head | — | — | — | — |
|
Table 1. Parameters of Yolo-C
Model | Backbone | FAB | FLOPs | Params/106 | |
---|
Yolov3 | DarkNet-53 | | 32.77 | 61.55 | | Ours | DarkNet-53 | √ | 35.58 | 64.65 | | DarkNet-C | √ | 61.85 | 105.93 | |
|
Table 2. Analysis of different network complexity
No. | Model | Backbone | FAB | AP for gun /% | AP for knife /% | AP for pliers /% | AP for wrench /% | AP for scissor /% | mAP /% | Detection rate /(frame·s-1) |
---|
1 | SSD | ResNet50 | | 85.69 | 72.44 | 51.21 | 60.62 | 41.99 | 62.39 | 56 | 2 | FASF | ResNet101 | | 82.77 | 70.23 | 48.6 | 57.7 | 38.7 | 59.6 | 48 | 3 | Yolov3 | DarkNet-53 | | 88.67 | 76.53 | 52.49 | 61.41 | 42.6 | 64.34 | 57 | 4 | Faster-RCNN | ResNet101 | | 93.83 | 83.74 | 58.85 | 71.58 | 54.1 | 72.18 | 10 | 5 | Ours | DarkNet-53 | √ | 90.1 | 79.57 | 55.67 | 64.81 | 51.00 | 68.23 | 55 | 6 | DarkNet-C | | 91.6 | 82.1 | 58.7 | 69.7 | 54.4 | 71.10 | 42 | 7 | DarkNet-C | √ | 93.14 | 83.12 | 60.18 | 73.82 | 58.14 | 73.68 | 40 |
|
Table 3. Ablation experiments on the Yolo-C network based on SIXray_OD dataset