An improved semi-supervised transfer learning method for infrared object detection neural network

Weipeng Li; Xiaogang Yang; Chuanxiang Li; Ruitao Lu; Pan Huang

doi:10.3788/IRLA20200511

Journals >Infrared and Laser Engineering >Volume 50 >Issue 3 >Page 20200511 > Article

Infrared and Laser Engineering
Vol. 50, Issue 3, 20200511 (2021)

An improved semi-supervised transfer learning method for infrared object detection neural network

Weipeng Li, Xiaogang Yang, Chuanxiang Li, Ruitao Lu, and Pan Huang

Author Affiliations

College of Missile Engineering, Rocket Force Engineering University, Xi’an 710025, China

show less

DOI: 10.3788/IRLA20200511 Cite this Article

Weipeng Li, Xiaogang Yang, Chuanxiang Li, Ruitao Lu, Pan Huang. An improved semi-supervised transfer learning method for infrared object detection neural network[J]. Infrared and Laser Engineering, 2021, 50(3): 20200511 Copy Citation Text

show less

Fig. 1. Influence of unlabeled samples on decision boundary

Download full size | View in the Article

Fig. 2. Procedures of semi-supervised transfer learning of infrared object detection neural network

Download full size | View in the Article

Fig. 3. IR features learned by object detection neural network

Download full size | View in the Article

Fig. 4. Comparison of object detection with only transfer learning and with semi-supervised transfer learning

Download full size | View in the Article

Algorithm 1: SSTL
Input: Detection neural network and the parameters , RGB dataset , labeled IR dataset , unlabeled IR dataset , weight of unsupervised loss , .
Output: Trained neural network
1. Initialize parameters of neural network , weight of unsupervised loss ;
2. Pre-train on RGB dataset ;
3. Adjust the number of output channels of classifier according to the number of categories of IR labels;
4. FOR Epoch t
5. 　 FOR Each Batch
6. 　　Sampling BatchSize training data form and ;
7. 　　Getting the predictions , of batch images through forward propagation;
8. 　　Calculate the loss of predictions for supervised samples with Eq.(1);
9. 　　For each , find the neighbourhood ;
10. 　　For each , find the neighbourhood ;
11. 　　Calculate the unsupervised loss of predictions through Eq.(7);
12. 　　Calculate the gradient for semi-supervised loss Eq.(2);
13. 　　Update ;
14. 　 END FOR
15. 　Update the weight of unsupervised loss α with Eq.(9);
16. END FOR

Table 1. [in Chinese]

View in the Article

Classification	Training		Test	Total
Classification	Labeled	Unlabeled	Test	Total
Launcher	36	33	38	107
Tank	15	19	19	53
Airplane	42	41	42	125
Battleship	51	55	51	157
Total	144	148	150	442

Table 2. [in Chinese]

View in the Article

Method		Epochs	Launcher	Tank	Airplane	Battleship	mAP
Supervised transfer learning	Faster R-CNN	60	0.946	0.995	0.965	0.980	0.972
Supervised transfer learning	YOLO-v3	80	0.964	0.848	0.919	0.979	0.927
Semi-supervised transfer learning	Faster R-CNN	60	0.971	0.997	0.962	1.000	0.983
Semi-supervised transfer learning	YOLO-v3	80	1.000	0.973	0.936	1.000	0.975

Table 3. [in Chinese]

Download Citation

Save the article for my favorites

Paper Information