Infrared Weak Target Detection Method Based on Cross Connection and Fusion Attention Mechanism

Feature map	Anchors before optimization	Optimized anchors
64×64	［10，13， 16，30， 33，23］	［4，4， 6，6 10，10］
32×32	［30，61， 62，45， 59，119］	［12，12， 20，13， 18，18］
16×16	［116，90， 156，198， 373，326］	［21，20， 30，30， 32，32］

Table 2. SCR values in experimental dataset images

View table

View in Article

Table 2. SCR values in experimental dataset images

Sequence	Minimum SCR	Maximum SCR	Total frame rate
1	0.000 2	2.513 3	200
2	0.001 3	2.729 8	200
3	0.002 7	3.512 1	200
4	0.000 3	0.639 9	201
5	0.000 3	1.322 1	249
6	0.001 8	1.507 2	188
7	0.005 6	2.484 4	191
8	0.001 7	1.899 7	713
9	0.005 4	1.740 7	188
10	0.001 6	2.247 9	250
11	0.006 6	8.287 9	200
12	0.002 0	1.256 0	200
13	0.000 2	1.802 6	250
14	0.001 1	1.737 0	250

Table 3. Results of ablation experiment

View table

View in Article

Table 3. Results of ablation experiment

Baseline	C3CSA	BIFPN	Detect	Anchors	NWD	P/%	R/%	AP/%	Parameters	Size/MB
√						95.9	95.8	96.8	7 012 822	14.3
√	√					96.1	95.2	96.7	7 349 176	15.0
√		√				97.7	96.2	97.7	7 161 311	14.6
√			√			95.5	94.9	97.1	5 375 510	11.1
√				√		97.8	97.8	97.7	7 012 822	14.3
√					√	97.1	97.5	97.3	7 012 822	14.3
√	√	√	√	√	√	97.8	98.7	98.4	5 749 507	11.9

Table 4. Comparison results of different algorithms
View table
View in Article
Table 4. Comparison results of different algorithms
Model AP/% Size/MB FPS
MDvsFA 81.5 15.8 15.1
AGPC 95.6 49.7 3.4
DNANet 91.8 19.0 10.9
OSCAR 57.7 202.7 16.7
IAANet 21.3 50.0 5.7
Faster-rcnn 68.1 1100.0 7.9
yolov7 89.9 74.8 57.5
Ours 98.4 11.9 107.0

Tools

Get Citation

Copy Citation Text

Hui LI, Zhengzhou LI, Yuxin YANG, Congyu HAO, Haitao LIU. Infrared Weak Target Detection Method Based on Cross Connection and Fusion Attention Mechanism[J]. Acta Photonica Sinica, 2024, 53(9): 0910002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites