基于YOLO v3的红外末制导典型目标检测 
下载： 1379次
Typical Target Detection for Infrared Homing Guidance Based on YOLO v3
火箭军工程大学作战保障学院, 陕西 西安 710025
图 & 表
图 1. 预测框和标定框的位置信息图
Fig. 1. Illustration of positions of predicting and bounding boxes
下载图片 查看原文
图 2. 机场可见光图像
Fig. 2. Visible image of airport
下载图片 查看原文
图 3. 机场红外图像
Fig. 3. Infrared image of airport
下载图片 查看原文
图 4. 动量法加速收敛
Fig. 4. Accelerated convergence by SGD with momentum
下载图片 查看原文
图 5. 基于改进YOLO v3的红外末制导目标检测方法示意图
Fig. 5. Diagram of infrared homing guidance target detection method based on improved YOLO v3
下载图片 查看原文
图 6. 平均交并比与锚点框的关系示意图
Fig. 6. Relationship between IOU and number of anchor box
下载图片 查看原文
图 7. 损失曲线
Fig. 7. Curve of loss
下载图片 查看原文
图 8. 交叉验证损失曲线
Fig. 8. Curve of cross validation loss
下载图片 查看原文
表 1对比实验参数表
Table1. Parameters for contrastive experiments
| Algorithm | Training epoch | Total training time | Parameter |
|---|
| SGD with momentum | 1-200 | 22 h 24 min | Lr0=10-6,decay is 0.5,decay period is 5,momentum is 0.999 | | Weighted SGDwith momentum | 1-200 | 22 h 32 min | Lr0=10-7,decay is 0.5,decay period is 5,momentum is 0.999 | | Adam | 1-200 | 23 h 20 min | Lr0=10-3,β1 = 0.9,β2 = 0.999,ε = 10-8 | | Weighted Adam | 1-200 | 23 h 28 min | Lr0=10-3,β1 = 0.9,β2 = 0.999,ε = 10-8 | | Adam in unitedtraining method | 1-100 | 11 h 40 min | Lr0=10-3,β1 = 0.9,β2 = 0.999,ε = 10-8 | | SGD with momentum inunited training method | 101-200 | 11 h 10 min | Lr0=10-4,decay is 0.5,decay period is 10,momentum is 0.999 | | Adam in weightedunited training method | 1-100 | 11 h 44 min | Lr0=10-3,β1 = 0.9,β2 = 0.999,ε = 10-8 | | SGD with momentumin weighted unitedtraining method | 101-200 | 11 h 16 min | Lr0=10-4,decay is 0.5,decay period is 10,momentum is 0.999 |
|
查看原文
表 2不同方法测试结果对比
Table2. Comparison of different testing methods
| Algorithm | SGD withmomentum | Adam | WeightedAdam | Unitedtraining | Weighted unitedtraining |
|---|
| Epoch | 100 | 200 | 100 | 200 | 100 | 200 | 100 | 200 | 100 | 200 | | mAP /% | - | 16.24 | 32.56 | 67.55 | 58.06 | 67.73 | 32.56 | 77.73 | 58.06 | 77.89 | | Special vehicle /% | - | 13.05 | 47.49 | 63.16 | 53.32 | 61.79 | 47.49 | 67.36 | 53.32 | 69.56 | | Ship /% | 5.30 | 18.23 | 21.97 | 93.54 | 87.09 | 86.85 | 21.97 | 98.48 | 87.09 | 97.97 | | Aircraft /% | - | 17.45 | 28.21 | 45.95 | 33.77 | 54.54 | 28.21 | 67.34 | 33.77 | 66.14 | | MA /% | 91.24 | 62.38 | 64.20 | 57.78 | 57.72 | | FA /% | 7.27 | 15.45 | 8.37 | 14.29 | 13.58 | | Rate /(frame·s-1) | 26.79 | 26.88 | 25.83 | 25.44 | 25.62 | 25.53 | 25.38 | 25.27 | 25.62 | 25.33 |
|
查看原文
陈铁明, 付光远, 李诗怡, 李源. 基于YOLO v3的红外末制导典型目标检测[J]. 激光与光电子学进展, 2019, 56(16): 161502. Tieming Chen, Guangyuan Fu, Shiyi Li, Yuan Li. Typical Target Detection for Infrared Homing Guidance Based on YOLO v3[J]. Laser & Optoelectronics Progress, 2019, 56(16): 161502.
PDF全文
