Recognition of Wild Animals Using Infrared Camera Images Based on YOLOv5

Minglun Yang; Xu Zhang; Ying Guo; Xinwen Yu; Yanan Hou; Jiajun Gao

doi:10.3788/LOP202259.1215015

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 12 >Page 1215015 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 12, 1215015 (2022)

Recognition of Wild Animals Using Infrared Camera Images Based on YOLOv5

Minglun Yang^1、2, Xu Zhang^1、2, Ying Guo^1、2、*, Xinwen Yu^1、2, Yanan Hou^1、2, and Jiajun Gao^1、2

Author Affiliations

¹Research Institute of Forest Resource Information Techniques, Chinese Academy of Forestry, Beijing 100091, China

²Key Laboratory of Forestry Remote Sensing and Information Technology, State Forestry and Grassland Administration, Beijing 100091, China

show less

DOI: 10.3788/LOP202259.1215015 Cite this Article Set citation alerts

Minglun Yang, Xu Zhang, Ying Guo, Xinwen Yu, Yanan Hou, Jiajun Gao. Recognition of Wild Animals Using Infrared Camera Images Based on YOLOv5[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215015 Copy Citation Text

show less

Fig. 1. Network structure diagram of YOLOv5

Download full size

Fig. 2. Complexity of image background

Download full size

Fig. 3. Trend of total loss of YOLOv5

Download full size

Fig. 4. Detection effects of YOLOv3-tiny and YOLOv5m. (a) YOLOv3-tiny; (b) YOLOv5m

Download full size

Fig. 5. Examples of detection results in complex background

Download full size

Residual componentmodel	YOLOv5s	YOLOv5m	YOLOv5l	YOLOv5x
CSP1_1_A	1	2	3	4
CSP1_3_B	3	6	9	12
CSP1_3_C	3	6	9	12
CSP2_1_A	1	2	3	4
CSP2_1_B	1	2	3	4
CSP2_1_C	1	2	3	4
CSP2_1_D	1	2	3	4
CSP2_1_E	1	2	3	4

Table 1. Depth comparison of four network structures

Filter model	YOLOv5s	YOLOv5m	YOLOv5l	YOLOv5x
Focus	32	48	64	80
CBL_A	64	96	128	160
CBL_B	128	192	256	320
CBL_C	256	384	512	640
CBL_D	512	768	1024	1280

Table 2. Width comparison of four network structures

Class	Train	Val	Test	Total
Sika deer	650	73	93	816
Tufted deer	538	54	63	655
Wild boar	571	74	72	717
Impala	567	71	69	707
Tragopan temminckii	401	32	40	473
Total	2727	304	337	3368

Table 3. Number of images in each class

Model	F1-score	mAP	Model size /MB	Inference time /ms
YOLOv5s	0.917	0.965	14	9.5
YOLOv5m	0.91	0.961	42	12.9
YOLOv5l	0.908	0.96	93	17.9
YOLOv5x	0.918	0.965	170	21.7

Table 4. Comparison of experimental results of each model on the test set

Class	YOLOv5s	YOLOv5m	YOLOv5l	YOLOv5x
Sika deer	3	6	2	2
Tufted deer	12	9	12	10
Boar	0	0	0	2
Impala	12	4	4	3
Tragopan temminckii	1	1	3	1
Empty prediction box	0	0	0	2
Total	28	20	21	21

Table 5. Statistics of number of false and missed detections

Model	F1-score	mAP	Weight /MB	Inference time /ms
YOLOv5m	0.91	0.961	42	12.9
YOLOv3-tiny	0.842	0.943	16.7	3.7

Table 6. Comparison of YOLOv5m and YOLOv3-tiny

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information