Optics and Precision Engineering, Volume. 32, Issue 8, 1227(2024)
Improving the lightweight VTG-YOLOv7-tiny for steel surface defect detection
Figures & Tables(18)
Fig. 8. Comparison of AP values of this paper's algorithm and the original model for various types of defects
Fig. 9. Comparison of the detection effect of this paper's algorithm and the original model
Table 1. Improved anchoring frame sizes for each detection layer
View tableView in ArticleTable 1. Improved anchoring frame sizes for each detection layer
特征图尺寸 感受野 NEU-DET数据集锚定框大小 Severstal数据集锚定框大小 10×10 超大 （361，266）（246，560）（490，540） （478，238）（204，502）（606，613） 20×20 大 （117，517）（217，355）（438，190） （428，122）（267，235）（112，576） 40×40 中 （51，529）（176，193）（524，77） （48，573）（159，187）（157，318） 80×80 小 （61，130）（135，111）（82，238） （60，121）（132，111）（76，239）
Table 2. Improved VoVGA-FPN comparison experiments
View tableView in ArticleTable 2. Improved VoVGA-FPN comparison experiments
Model mAP% Params/M FLOPs/G FPS/ms PANet 68.7 6.02 13.1 108 AFPN 70.7 7.11 14.1 108 Ours 71.2 6.35 12.7 111
Table 3. Experiments with different positions of the TCA module
View tableView in ArticleTable 3. Experiments with different positions of the TCA module
Location mAP% Params/M FLOPs/G Baseline 68.7 6.020 1 13.1 情况1 70.1 6.020 4 13.2 情况2 70.8 6.020 5 13.2 情况3 69.8 6.020 8 13.2
Table 4. TCA comparison experiments
View tableView in ArticleTable 4. TCA comparison experiments
Model mAP% Params/M FLOPs/G Baseline 68.7 6.020 1 13.1 SE 70.4 11.570 30.8 GC 70.1 6.037 3 13.2 CA 68.3 6.034 5 13.5 CBAM 70.1 6.028 6 13.2 TA 69.9 6.020 5 13.2 TCA（Ours） 70.8 6.020 5 13.2
Table 5. Experiments with different positions of the GSConv module
View tableView in ArticleTable 5. Experiments with different positions of the GSConv module
Location mAP% Params/M FLOPs/G FPS/ms A 68.7 6.02 13.1 108 B 70.1 4.82 9.9 94 C 70.3 4.31 10.3 121 D 72.0 3.12 7.1 100
Table 6. New detection layer with Neck with and without convolution experiments
View tableView in ArticleTable 6. New detection layer with Neck with and without convolution experiments
Model mAP% Params/M FLOPs/G FPS/ms A 68.7 6.02 13.1 108 B 72.1 7.87 13.5 111 C 73.0 10.2 13.9 80
Table 7. Ablation experiments on NEU-DET and Severstal datasets
View tableView in ArticleTable 7. Ablation experiments on NEU-DET and Severstal datasets
Dataset Method mAP% Params/M FLOPs/G FPS/ms P% R% NEU-DET M1 68.7 6.02 13.1 108 61.4 72.7 M2 70.8 6.02 13.2 119 63.8 73.7 M3 72.5 7.87 13.5 113 66.7 74.2 M4 73.6 7.04 12.7 76 67.1 72.0 M5 74.4 5.41 8.9 87 68.5 74.5 Severstal M1 65.6 6.02 13.1 61 58.1 65.2 M2 67.6 6.02 13.2 79 65.8 70.2 M3 71.5 7.87 13.5 70 64.5 72.1 M4 72.7 7.04 12.7 52 66.8 67.7 M5 74.1 5.41 8.9 78 67.0 72.2
Table 8. Comparison of experiments before and after enrichment of the Severstal dataset
View tableView in ArticleTable 8. Comparison of experiments before and after enrichment of the Severstal dataset
Dataset Model mAP% Params/M FLOPs/G FPS/ms SeverstalA M1 71.5 6.02 13.1 89 M5 75.2 5.41 8.9 94 SeverstalB M1 65.6 6.02 13.1 61 M5 74.1 5.41 8.9 78
Table 9. Experiments comparing NEU-DET and Severstal datasets
View tableView in ArticleTable 9. Experiments comparing NEU-DET and Severstal datasets
Dataset Model mAP0.5/% Params/ M FLOPs/G Size/MB NEU-DET Faster R-CNN 65.7 72.0 167.3 108.0 SSD 61.0 41.1 145.3 93.1 YOLOv3 67.0 61.5 155.0 117.0 YOLOv3-tiny 46.5 8.67 12.9 16.6 YOLOv4 51.0 52.5 119.8 102.6 YOLOv4-tiny 54.6 5.90 16.18 22.5 YOLOv5s 70.1 7.07 16.4 13.7 YOLOX-s 71.8 8.0 21.6 16.3 YOLOv7 70.0 37.2 104.8 74.8 YOLOv7-tiny 68.7 6.02 13.1 11.7 YOLOv8s 72.1 11.1 28.4 21.4 文献［ 1 ］74.1 23.9 — — 文献［ 16 ］82.4 15.37 — 29.7 文献［ 18 ］77.2 7.03 — — VTG-YOLOv7-tiny（Ours） 74.4 5.41 8.9 10.8 Severstal YOLOv3-tiny 56.4 8.67 12.9 16.6 YOLOv4-tiny 59.6 5.9 16.18 22.5 YOLOv5s 72.4 7.07 16.4 13.7 YOLOX-s 73.8 8.0 21.6 16.3 YOLOv7-tiny 65.6 6.02 13.1 11.7 YOLOv8s 71.8 11.1 28.4 21.4 VTG-YOLOv7-tiny（Ours） 74.1 5.41 8.9 10.8
Tools
Get Citation
Copy Citation Text
Liming LIANG, Pengwei LONG, Yao FENG, Baohe LU. Improving the lightweight VTG-YOLOv7-tiny for steel surface defect detection[J]. Optics and Precision Engineering, 2024, 32(8): 1227
Paper Information
Category:
Received: Oct. 16, 2023
Accepted: --
Published Online: May. 29, 2024
The Author Email: LONG Pengwei (2637018663@qq.com)
© Copyright 2018-2021 | Chinese Laser Press.
All Rights Reserved 沪ICP备15018463号-20