Author Affiliations
Engineering Research Center for Metallurgical Automation and Measurement Technology, Ministry of Education, Wuhan University of Science and Technology, Wuhan 430081, Hubei , Chinashow less
Fig. 1. Diagram of YOLOv4 network structure
Fig. 2. Prediction box of YOLOv4 algorithm in 19×19 cells
Fig. 3. Diagram of improved YOLOv4 network structure
Fig. 4. CSPNet structured in SPP module
Fig. 5. CSPNet structured in continuous convolution module
Fig. 6. Depthwise separable convolution
Fig. 7. Scatter plot of the size distribution of sample ground truth box and prior box
Fig. 8. Accuracy comparison between K-means algorithm and K-means++ algorithm under original YOLOv4
Fig. 9. Comparison of detection results before and after algorithm improvement
Fig. 10. Comparison of detection results on each class before and after algorithm improvement
Method | Neck module | Parameters / | mAP /% | Speed /(frame·s-1) |
---|
1 | SPP+PAN | 5.35 | 81.0 | 66.1 | 2 | SPP+CSPPAN | 4.65 | 81.8 | 66.5 | 3 | CSPSPP+PAN | 5.95 | 82.6 | 63.6 | 4 | CSPSPP+CSPPAN | 5.25 | 83.6 | 62.3 |
|
Table 1. Performance results of the neck network integrated into the CSPNet structure
Method | Backbone | Neck | Parameters/ | mAP /% | Speed /(frame·s-1) |
---|
1 | CSPDarknet53 | PAN | 5.95 | 82.6 | 63.5 | 2 | CSPDarknet53 | CSPPAN | 5.25 | 83.6 | 62.3 | 3 | DS-CSPDarknet53 | PAN | 4.41 | 81.9 | 67.0 | 4 | DS-CSPDarknet53 | CSPPAN | 3.89 | 83.2 | 68.0 | 5 | CSPDarknet53 | DS-PAN | 4.72 | 82.0 | 68.7 | 6 | CSPDarknet53 | DS-CSPPAN | 4.60 | 82.9 | 65.8 | 7 | DS-CSPDarknet53 | DS-PAN | 3.28 | 81.9 | 72.9 | 8 | DS-CSPDarknet53 | DS-CSPPAN | 3.24 | 83.0 | 72.1 |
|
Table 2. Impact of depthwise separable convolution on network performance
Improvement strategy | Weight size /MB | mAP /% | Speed /(frame·s-1) |
---|
K-means++ | CSP-Neck | DS |
---|
| | | 102 | 79.8 | 66.1 | | | √ | 52 | 79.3 | 78.6 | | √ | | 101 | 81.2 | 62.4 | | √ | √ | 65 | 80.6 | 71.9 | √ | | | 102 | 81.0 | 66.1 | √ | | √ | 52 | 80.8 | 78.1 | √ | √ | | 101 | 83.6 | 62.3 | √ | √ | √ | 65 | 83.0 | 72.1 |
|
Table 3. Performance comparison of different improvements
Model | Weight size /MB | mAP /% | Speed /(frame·s-1) |
---|
Faster-RCNN-resnet50 | 315 | 79.0 | 21.3 | SSD-resnet50 | 108 | 78.2 | 40.5 | YOLOv3 | 118 | 75.2 | 60.8 | YOLOv4 | 102 | 79.8 | 66.1 | YOLOv5 | 94 | 81.5 | 72.5 | Improved YOLOv4 | 65 | 83.0 | 72.1 |
|
Table 4. Performance comparison of different detection algorithms