Chinese Optics, Volume. 16, Issue 5, 1045(2023)
Lightweight YOLOv5s vehicle infrared image target detection
Fig. 3. (a) Ordinary convolution and (b) Ghost convolution (Φ is a linear operation)
Fig. 6. Data augmentation results. (a) Mosaic augmentation; (b) MixUp augmentation; (c) Copy-Paste augmentation
Fig. 7. Detection results of different algorithms. (a) YOLOv3-tiny; (b) YOLOv4-tiny; (c) YOLOv5n; (d) YOLOv6-N; (e) YOLO7-tiny; (f) YOLO5s; (g) proposed in this paper
Table 1. Optimized prior anchor size
View tableView in ArticleTable 1. Optimized prior anchor size
特征图尺度 160×160 80×80 40×40 感受野大小 小 中 大 [6,8] [14,37] [35,94] 先验框 [7,19] [31,26] [96,68] [15,13] [50,37] [154,145]
Table 2. Performance comparison of lightweight for YOLOv5s and YOLOv5s-G
View tableView in ArticleTable 2. Performance comparison of lightweight for YOLOv5s and YOLOv5s-G
Model t/hours Size/MB Params/M GFLOPs P(%) R(%) mAP(%) FPS YOLOv5s 48.77 13.70 7.02 15.8 87.1 69.8 80.8 119 YOLOv5s-G 30.25 7.46 3.68 8.0 86.1 66.3 77.5 137
Table 3. Performance comparison of different loss functions
View tableView in ArticleTable 3. Performance comparison of different loss functions
Model t/hours P(%) R(%) mAP(%) FPS YOLOv5s-G 30.25 86.1 66.3 77.5 137 YOLOv5s-G-EIoU 24.31 84.5 68.7 78.9 141 YOLOv5s-G-SIoU 24.62 85.8 67.2 77.8 139 YOLOv5s-G-αIoU 23.50 85.9 69.3 79.8 147
Table 4. Performance comparison of multi-scale fusion
View tableView in ArticleTable 4. Performance comparison of multi-scale fusion
Model t/hours Size/MB Params/M GFLOPs P(%) R(%) mAP(%) FPS YOLOv5s-G-αIoU 23.50 7.46 3.68 8.0 85.9 69.3 79.8 147 YOLOv5s-G1-αIoU 26.89 8.60 3.75 9.6 86.0 73.6 83.6 125 YOLOv5s-G2-αIoU 24.56 2.73 0.95 7.2 84.5 72.8 82.9 154 YOLOv5s-G2-αIoU-KMeans 25.62 2.73 0.95 7.2 85.5 72.4 83 154
Table 5. Performance comparison of different attention mechanisms
View tableView in ArticleTable 5. Performance comparison of different attention mechanisms
Model t/hours Size/MB Params/M GFLOPs P(%) R(%) mAP(%) FPS YOLOv5s-G2-αIoU-KMeans 25.62 2.73 0.95 7.2 85.5 72.4 83 154 YOLOv5s-G2-αIoU-KMeans-SE 30.95 2.75 0.96 7.2 86.0 73.5 84.1 149 YOLOv5s-G2-αIoU-KMeans-ECA 26.06 2.73 0.95 7.2 85.5 73.8 84.2 145 YOLOv5s-G2-αIoU-KMeans-CBAM 28.21 2.76 0.96 7.3 85.7 73.4 84 135 YOLOv5s-G2-αIoU-KMeans-CA 28.62 2.76 0.96 7.3 86.6 73.6 84.3 139
Table 6. SPD-Conv effect
View tableView in ArticleTable 6. SPD-Conv effect
Model t/hours Size/MB Params/M GFLOPs P(%) R(%) mAP(%) FPS YOLOv5s-G2-αIoU-Kmeans-CA 28.62 2.76 0.96 7.3 86.6 73.6 84.3 139 YOLOv5s-G2-αIoU-Kmeans-CA-SPD 30.28 3.0 1.09 9.4 87.4 74.6 85.0 132
Table 7. Comparison with other advanced algorithms
View tableView in ArticleTable 7. Comparison with other advanced algorithms
Model Size/MB Params/M GFLOPs P(%) R(%) mAP(%) FPS SSD 186.0 23.70 115.7 68.9 55.7 63.2 88 EfficientDet 302.0 39.40 107.5 72.8 58.4 67.8 52 YOLOv4+GhostNet 150.3 39.30 25.6 81.1 66.9 77.7 112 YOLOv5-MobileNetV3 7.9 4.0 9.3 83.7 67.5 76.9 128 YOLOv3-tiny 16.6 8.67 12.9 79.3 54.9 62.9 175 YOLOv4-tiny 12.9 6.27 16.2 78.9 57.3 67.2 149 YOLOv5n 3.7 1.76 5.1 83.6 66.1 76.6 164 YOLOv6-N 9.3 4.30 11.1 84.8 71.5 80.3 208 YOLOv7-tiny 12.3 6.02 13.2 84.2 74.7 83.6 143 YOLOv5s 13.7 7.02 15.8 87.1 69.8 80.8 119 proposed in this paper 3.0 1.09 9.4 87.4 74.6 85.0 132
Get Citation
Copy Citation Text
Yan-lei LIU, Meng-zhe LI, Xuan-xuan WANG. Lightweight YOLOv5s vehicle infrared image target detection[J]. Chinese Optics, 2023, 16(5): 1045
Category: Original Article
Received: Dec. 14, 2022
Accepted: Mar. 24, 2023
Published Online: Oct. 27, 2023
The Author Email:
