Optics and Precision Engineering, Volume. 32, Issue 2, 301(2024)
Flue-cured tobacco leaf grade detection through multi-receptive field features fusing adaptively and dynamic loss adjustment
Figures & Tables(17)
Fig. 11. Visualization of FTGDNet and YOLOv5 feature extraction process
Table 1. Backbone network comparison study
View tableView in ArticleTable 1. Backbone network comparison study
Module FLOPs/G Parameters/M mAP/% Inference time/ms Val Test ShuffleNet 9.5 5.0 56.0 51.8 6.0 GhostNet 11.2 5.7 59.5 55.5 6.3 CSPNet 16.4 7.1 78.8 75.7 7.0 CSPNet+ShuffleNet 18.1 7.7 77.3 77.0 9.0 CSPNet+GhostNet 19.4 8.0 83.5 81.2 9.3
Table 2. Module ablation study
View tableView in ArticleTable 2. Module ablation study
Module EVCB MRFA_d MCIoU_Loss FLOPs
/G
Parameters
/M
mAP/% Inference time/ms Val Test Baseline × × × 19.4 8.0 83.5 81.2 9.3 Baseline+EVCB √ × × 21.5 10.8 86.6 82.2 10.9 Baseline+EVCB+MRFA_d √ √ × 22.9 12.4 89.2 85.3 12.6 Baseline+EVCB+MRFA_d+MCIoU_Loss √ √ √ 22.9 12.4 90.0 87.4 12.6
Table 3. Effect comparison of bottleneck module
View tableView in ArticleTable 3. Effect comparison of bottleneck module
Module FLOPs/G Parameters/M mAP/% Inference time/ms Val Test Bottleneck 19.5 8.2 81.5 81.4 9.4 BottleneckCSP 19.4 8.1 82.6 80.6 9.4 C3 19.4 8.0 83.5 81.2 9.3 EVCB 21.5 10.8 86.6 82.2 10.9
Table 4. Performance comparison of four MRFA structures
View tableView in ArticleTable 4. Performance comparison of four MRFA structures
Module FLOPs/G Parameters/M mAP/% Inference time/ms Val Test MRFA_a 22.8 12.4 87.1 84.4 12.1 MRFA_b 22.8 12.4 88.6 83.7 12.3 MRFA_c 22.8 12.4 88.8 85.0 12.3 MRFA_d 22.9 12.4 89.2 85.3 12.6
Table 5. Performance comparison of different Loss functions
View tableView in ArticleTable 5. Performance comparison of different Loss functions
Loss mAP/% Val Test GIoU_Loss 85.9 82.4 DIoU_Loss 87.7 84.8 SIoU_Loss 88.2 85.3 CIoU_Loss 89.2 85.3 MCIoU_Loss 90.0 87.4
Table 6. Comparison of model effects
View tableView in ArticleTable 6. Comparison of model effects
Module FLOPs/G Parameters/M mAP/% Inference time/ms Val Test TOOD 181.0 31.8 69.2 68.5 63.3 Faster R-CNN 206.7 41.2 72.1 71.3 62.5 Sparse R-CNN 150.0 106.0 77.4 74.7 58.5 Double_head R-CNN 480.9 46.8 79.1 78.2 122.0 Dynamic R-CNN 206.7 41.2 82.6 80.9 64.9 SOBL 413.3 71.3 86.2 85.7 204.1 YOLOx-Tiny 19.0 5.4 62.3 61.2 13.1 YOLOx 33.3 8.9 81.7 78.2 24.6 YOLOv3-Tiny 12.9 8.6 50.8 49.1 3.4 YOLOv3 154.9 61.5 82.4 79.7 27.7 YOLOr_CSP 119.2 92.5 74.7 68.2 13.4 YOLOr_CSPx 225.3 96.4 74.5 67.1 25.4 YOLOv5n 4.2 1.8 77.7 73.1 5.4 YOLOv5s 16.4 7.1 78.8 75.7 7.0 YOLOv5m 48.1 20.9 84.1 83.0 13.0 YOLOv6n 11.4 4.7 77.9 77.0 3.6 YOLOv6s 30.5 13.1 88.0 85.7 12.7 YOLOv6m 72.2 28.3 89.5 85.8 15.7 YOLOv7-Tiny 13.2 6.0 86.9 83.9 4.7 YOLOv7 103.5 36.5 88.2 84.2 12.8 FTGDNet 22.9 12.4 90.0 87.4 12.6
Tools
Get Citation
Copy Citation Text
Zifen HE, Yang LUO, Yinhui ZHANG, Guangchen CHEN, Dongdong CHEN, Lin XU. Flue-cured tobacco leaf grade detection through multi-receptive field features fusing adaptively and dynamic loss adjustment[J]. Optics and Precision Engineering, 2024, 32(2): 301
Paper Information
Category:
Received: May. 19, 2023
Accepted: --
Published Online: Apr. 2, 2024
The Author Email: ZHANG Yinhui (zyhhzf1998@163.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20