GPNet：Lightweight infrared image target detection algorithm

Fig. 11. Detection comparison of GPNet and SOTA algorithms on the FLIR IR test set(a)YOLOv4 detection results,（b）YOLOv5-m detection results,（c）YOLOX-m detection results,（d）GPNet（ours）detection results

Download full size

View in Article

Table 1. Quantitative comparison of GPNet and SOTA algorithms on the FLIR IR test set

View table

View in Article

Table 1. Quantitative comparison of GPNet and SOTA algorithms on the FLIR IR test set

Model	AP/（%）		mAP50/（%）	Recall/（%）		F1		Params/M	FLOPs/G	Time/ms
Model	person	car	mAP50/（%）	person	car	person	car	Params/M	FLOPs/G	Time/ms
FasterR-CNN	39.09	61.67	50.38	47.06	69.84	0.43	0.47	136.7	252.7	75
SSD	43.78	58.72	51.25	20.34	42.60	0.33	0.58	23.7	115.7	15
YOLOv3	73.73	85.93	79.83	59.36	77.89	0.70	0.81	61.5	65.5	19
YOLOv4	78.13	84.74	81.44	61.45	73.99	0.72	0.80	63.9	59.8	25
YOLOv5-m	75.24	85.79	80.52	54.25	74.20	0.68	0.81	21.1	21.3	17
YOLOX-m	72.02	80.46	76.24	52.43	68.16	0.66	0.77	25.3	31.1	16
YOLOv4+GhostNet	69.41	86.14	77.77	49.17	76.00	0.63	0.81	39.3	25.6	17
GPNet（本文）	72.65	84.83	78.74	47.32	71.95	0.62	0.79	12.3	7.2	14

Table 2. Quantitative comparison of GPNet and SOTA algorithms on the KAIST IR test set

View table

View in Article

Table 2. Quantitative comparison of GPNet and SOTA algorithms on the KAIST IR test set

Model	Size	AP/（%）	Recall/（%）	F1
Faster R-CNN YOLOv4	416×416 416×416	39.52 50.45	55.49 49.49	0.40 0.54
YOLOv5-m	416×416	50.69	44.42	0.54
YOLOv5-s	416×416	50.18	44.65	0.53
YOLOX-m	416×416	54.41	48.82	0.56
YOLOX-s	416×416	53.49	47.27	0.55
GPNet（本文）	416×416	55.04	47.36	0.57

Table 3. Quantitative comparison of GPNet and SOTA algorithms on the CVC-09 IR test set

View table

View in Article

Table 3. Quantitative comparison of GPNet and SOTA algorithms on the CVC-09 IR test set

Model	Size	AP/（%）		mAP50/（%）	Recall/（%）		F1
Model	Size	person	car	mAP50/（%）	person	car	person	car
Faster R-CNN	416×416	42.39	67.92	55.15	52.40	75.42	0.40	0.54
YOLOv4	416×416	73.53	79.31	76.42	74.48	70.47	0.70	0.76
YOLOv5-m	416×416	75.31	82.07	78.69	80.89	79.22	0.72	0.76
YOLOv5-s	416×416	76.29	80.53	78.41	71.90	75.56	0.73	0.75
YOLOX-m	416×416	71.86	79.16	75.51	74.41	75.08	0.71	0.75
YOLOX-s	416×416	71.99	75.44	73.72	69.20	70.69	0.70	0.71
GPNet（本文）	416×416	76.29	86.51	81.40	70.59	85.23	0.75	0.84

Table 4. Quantitative comparison of GPNet and SOTA algorithms on the self-made campus infrared dataset

View table

View in Article

Table 4. Quantitative comparison of GPNet and SOTA algorithms on the self-made campus infrared dataset

Model	Size	AP/（%）	Recall/（%）	F1
Faster R-CNN YOLOv4	416×416 416×416	45.28 81.23	61.98 81.11	0.41 0.76
YOLOv5-m	416×416	79.63	78.25	0.76
YOLOv5-s	416×416	75.08	62.04	0.73
YOLOX-m	416×416	80.38	72.63	0.77
YOLOX-s	416×416	79.00	69.77	0.76
GPNet（本文）	416×416	81.46	78.35	0.80

Table 5. Ablation experiments in FLIR infrared dataset

View table

View in Article

Table 5. Ablation experiments in FLIR infrared dataset

Backbone	3-C	5-C	Dsample	Head	AP/（%）	mAP50/（%）	Recall/（%）	Params/M	Weight/MB
$□$ $□$		$□$			69.41 71.21	77.77 78.13	46.20 46.41	39.3 26.2	150.3 100.4
$□$	$□$	$□$			69.73	76.72	49.27	18.2	68.4
$□$	$□$	$□$		$□$	72.65	78.74	47.32	12.7	47.4
$□$	$□$	$□$	$□$	$□$	67.37	76.27	44.75	11.4	42.5

Tools

Get Citation

Copy Citation Text

Xian-Guo LI, Ming-Teng CAO, Bin LI, Yi LIU, Chang-Yun MIAO. GPNet：Lightweight infrared image target detection algorithm[J]. Journal of Infrared and Millimeter Waves, 2022, 41(6): 1092

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category: Research Articles

Received: May. 25, 2022

Accepted: --

Published Online: Feb. 6, 2023

The Author Email: Xian-Guo LI (lixianguo@tiangong.edu.cn)

DOI:10.11972/j.issn.1001-9014.2022.06.019

Topics

laser devices and laser physics

Lasers and Laser Optics

Laser physics

laser manufacturing

Instrumentation, Measurement and Metrology