Author Affiliations
School of Automation, Xi’an University of Posts & Telecommunications, Xi’an 710100, Shaanxi , Chinashow less
Fig. 1. Structure of the YOLOv5s network
Fig. 2. Structure of the SE module
Fig. 3. Structure of the SA module
Fig. 4. Structure of the YOLOv5s-AFA network
Fig. 5. Structure of the ERF module
Fig. 6. Structure of the iECA module
Fig. 7. Attention structure of the ASFF module
Fig. 8. X-ray security image dataset. (a)‒(f) Image 1‒ image 6
Fig. 9. Label distribution of training set. (a) Center point distribution; (b) width and height distribution
Fig. 10. Loss and mAP of five networks. (a) Loss curve; (b) mAP curve
Fig. 11. Comparison of detection results for X-ray images. (a) YOLOv4; (b) PP-YOLOv2; (c) YOLOv5x; (d) YOLOv5s; (e) YOLOv5s-AFA
Network | YOLOv5s | YOLOv5m | YOLOv5l | YOLOv5x |
---|
Model size /MB | 27 | 84 | 192 | 367 | Focus | 32 | 48 | 64 | 80 | CBL-1 | 64 | 96 | 128 | 160 | CBL-2 | 128 | 192 | 256 | 320 | CBL-3 | 256 | 384 | 512 | 640 | CBL-4 | 512 | 768 | 1024 | 1208 |
|
Table 1. Number of convolution kernels of Backbone in different YOLOv5 networks
Module | mAP /% | (mAP,0.50∶0.95)/% | NP |
---|
YOLOv5s | 87.2 | 55.9 | 0 | YOLOv5s+SA | 88.7 | 56.3 | 2048 | YOLOv5s+SAd | 88.9 | 56.4 | 940 | YOLOv5s +E-SAd | 90.4 | 57.6 | 17832 |
|
Table 2. Comparison results of SA modules
Module | mAP /% | (mAP,0.5∶0.95)/% | NP |
---|
YOLOv5s | 87.2 | 55.9 | 0 | YOLOv5s+SE | 88.1 | 56.3 | 43008 | YOLOv5s+ECA | 89.4 | 58.7 | 22868 | YOLOv5s+iECA(k=3) | 91.5 | 59.5 | 25668 | YOLOv5s+iECA(k=5) | 91.4 | 59.2 | 32786 |
|
Table 3. Comparison results of channel attention module
No. | Modules | ASFF | E-SAd | iECA | mAP /% |
---|
1 | YOLOv5s | × | × | × | 87.2 | 2 | YOLOv5s+E-SAd+iECA | × | √ | √ | 92.5 | 3 | YOLOv5s+ASFF | √ | × | × | 91.3 | 4 | YOLOv5s+ASFF+E-SAd | √ | √ | × | 92.2 | 5 | YOLOv5s+ASFF+iECA | √ | × | √ | 92.7 | 6 | YOLOv5s+ASFF+E-SAd+iECA | √ | √ | √ | 94.5 |
|
Table 4. Comparison results of ablation experiments of each module
Network | Backbone | P | R | mAP /% | Module size /m |
---|
Faster RCNN | VGG | 0.874 | 0.759 | 86.8 | 160 | RetinaNet | ResNet+FPN | 0.904 | 0.790 | 90.0 | 140 | YOLOv4 | CSP DarkNet53 | 0.920 | 0.812 | 91.7 | 240 | PP-YOLOv2 | ResNet50-vd | 0.949 | 0.865 | 93.4 | 83 | YOLOv5s | CSP DarkNet | 0.890 | 0.782 | 87.2 | 22 | YOLOv5x | CSP DarkNet | 0.968 | 0.889 | 95.6 | 320 | B-YOLO | Dark Net+CSP | 0.907 | 0.800 | 90.4 | 39 | YOLOv5+ | GhostBottleneck | 0.896 | 0.795 | 88.2 | 24 | YOLOv5s-AFA | CSP DarkNet | 0.967 | 0.871 | 94.5 | 26 |
|
Table 5. Detection results of different networks