Aerial small target detection algorithm based on context collaborative perception

Luxia YANG; Zekai LIU; Hongrui ZHANG; Yongjie MA

doi:10.37188/CJLCD.2025-0095

Chinese Journal of Liquid Crystals and Displays, Volume. 40, Issue 9, 1308(2025)

Aerial small target detection algorithm based on context collaborative perception

Luxia YANG^1,2, Zekai LIU^1,2, Hongrui ZHANG^1,2、*, and Yongjie MA³

Author Affiliations

¹School of Computer Science and Technology， Taiyuan Normal University， Jinzhong 030619， China

²Shanxi Provincial Key Laboratory of Intelligent Optimization Computing and Blockchain Technology， Jinzhong 030619， China

³School of Physics and Electronic Engineering， Northwest Normal University， Lanzhou 730070， China

show less

Abstract Get PDF(in Chinese)

Figures & Tables(19)

Fig. 1. Lightweight detection algorithm for context multi-scale collaborative sensing

Download full size

View in Article

Fig. 2. Comparison of C2f and LMEM structures

Download full size

View in Article

Fig. 3. MKFES module structure

Download full size

View in Article

Fig. 4. Structure diagram of CCFFAM module

Download full size

View in Article

Fig. 5. Diagram of RFAConv structure

Download full size

View in Article

Fig. 6. Dysample flow chart

Download full size

View in Article

Fig. 7. Detailed implementation of the Dysample module

Download full size

View in Article

Fig. 8. Design of small target detection head

Download full size

View in Article

Fig. 9. Heat maps of different models under two datasets

Download full size

View in Article

Fig. 10. Comparison of small target detection for different algorithms on VisDrone2019 dataset

Download full size

View in Article

Table 1. Parameter settings
View table
View in Article
Table 1. Parameter settings
参数值
imgsz 640
batch 4
workers 2
optimizer SGD
amp True
Ir0 0.01
lrf 0.01
momentum 0.937

Table 2. Influences of LMEM module on the model in different places
View table
View in Article
Table 2. Influences of LMEM module on the model in different places
位置 P R mAP50 mAP（50~95） Parameters/M
P5 0.44 0.33 0.334 0.193 2.96
P3、P4、P5 0.451 0.333 0.336 0.195 2.90
P2、P3、P4、P5 0.458 0.335 0.338 0.196 2.87

Table 3. Influence of downsampling convolution on the model in CCFFAM module
View table
View in Article
Table 3. Influence of downsampling convolution on the model in CCFFAM module
下采样卷积 P R mAP50 mAP（50~95） Parameters/M
Adown 0.469 0.367 0.373 0.222 2.13
Conv 0.484 0.365 0.38 0.227 2.17
RFAConv 0.484 0.370 0.385 0.230 2.17

Table 4. Influences of different position detection heads on the model
View table
View in Article
Table 4. Influences of different position detection heads on the model
检测头 P R mAP50 mAP（50~95） Parameters/M
P3、P4、P5 0.451 0.335 0.333 0.195 3.01
P2、P3、P4 0.476 0.353 0.369 0.22 2.01
P2、P3、P4、P5 0.457 0.36 0.364 0.218 2.92

Table 5. Influences of different loss functions on the model
View table
View in Article
Table 5. Influences of different loss functions on the model
损失函数 P R mAP50 mAP（50~95）
Inner-EIoU 0.475 0.373 0.376 0.225
Inner-DIoU 0.474 0.364 0.374 0.224
Focaler-MPDIoU 0.475 0.372 0.380 0.226
Focaler-EIoU 0.483 0.367 0.378 0.225
Focaler-CIoU 0.487 0.372 0.386 0.231

Table 6. Comparison of detection results of various small targets on VisDrone2019 dataset

View table

View in Article

Table 6. Comparison of detection results of various small targets on VisDrone2019 dataset

Class	YOLOv8n				Ours
Class	P	R	Map50	mAP（50~95）	P	R	mAP50	mAP（50~95）
all	0.451	0.335	0.333	0.195	0.487	0.372	0.386	0.231
pedestrian	0.456	0.346	0.335	0.155	0.544	0.413	0.454	0.212
people	0.525	0.235	0.282	0.101	0.573	0.322	0.375	0.152
bicycle	0.295	0.086	0.087	0.034	0.247	0.157	0.12	0.049
car	0.644	0.755	0.76	0.523	0.665	0.807	0.812	0.573
van	0.504	0.374	0.388	0.268	0.498	0.426	0.432	0.307
truck	0.396	0.304	0.278	0.181	0.469	0.3	0.319	0.215
tricycle	0.362	0.241	0.214	0.115	0.452	0.23	0.243	0.14
awning-tricycle	0.249	0.162	0.117	0.073	0.302	0.154	0.143	0.094
bus	0.624	0.449	0.482	0.34	0.586	0.456	0.505	0.359
motor	0.453	0.395	0.37	0.155	0.537	0.453	0.46	0.206

Table 7. Ablation experiments on VisDrone2019 dataset

View table

View in Article

Table 7. Ablation experiments on VisDrone2019 dataset

实验方法	LMEM模块	CCFFAM模块	改进检测头	Focaler-CIoU	P	R	mAP50	mAP（50~95）	Parameters/M
YOLOv8n					0.451	0.335	0.333	0.195	3.01
A	√				0.458	0.328	0.338	0.196	2.87
B		√			0.443	0.344	0.342	0.201	3.01
C			√		0.476	0.353	0.369	0.22	2.01
D				√	0.448	0.339	0.335	0.197	3.01
E		√	√		0.481	0.373	0.381	0.227	2.23
F	√	√	√		0.484	0.370	0.385	0.230	2.17
G	√	√	√	√	0.487	0.372	0.386	0.231	2.17

Table 8. Ablation experiments on DOTAv1 dataset

View table

View in Article

Table 8. Ablation experiments on DOTAv1 dataset

实验方法	LMEM模块	CCFFAM模块	改进检测头	Focaler-IoU	P	R	mAP50	mAP（50~95）	Parameters/M
YOLOv8n					0.637	0.391	0.415	0.25	3.01
A	√				0.659	0.387	0.417	0.25	2.80
B		√			0.641	0.394	0.420	0.253	3.02
C			√		0.653	0.396	0.425	0.254	2.10
D				√	0.647	0.391	0.417	0.251	3.01
E		√	√		0.651	0.4	0.427	0.256	2.34
F	√	√	√		0.657	0.398	0.429	0.256	2.17
G	√	√	√	√	0.663	0.403	0.429	0.257	2.17

Table 9. Comparison experiments on VisDrone2019 dataset

View table

View in Article

Table 9. Comparison experiments on VisDrone2019 dataset

模型	P	R	mAP50	mAP（50~95）	Parameters/M
YOLOv3n	0.447	0.333	0.329	0.189	4.06
YOLOv5n	0.442	0.337	0.332	0.181	2.50
YOLOv6n	0.404	0.312	0.294	0.167	4.23
YOLOv8n	0.451	0.335	0.333	0.195	3.01
ASF-YOLO	0.436	0.341	0.33	0.186	2.52
ours	0.487	0.372	0.386	0.231	2.17

Tools

Get Citation

Copy Citation Text

Luxia YANG, Zekai LIU, Hongrui ZHANG, Yongjie MA. Aerial small target detection algorithm based on context collaborative perception[J]. Chinese Journal of Liquid Crystals and Displays, 2025, 40(9): 1308

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites

Paper Information

Category:

Received: Apr. 25, 2025

Accepted: --

Published Online: Sep. 25, 2025

The Author Email: Hongrui ZHANG (zhanghongrui@tynu.edu.cn)

DOI:10.37188/CJLCD.2025-0095

Topics

Table 1. Parameter settings

Table 1. Parameter settings

Table 2. Influences of LMEM module on the model in different places

Table 2. Influences of LMEM module on the model in different places

Table 3. Influence of downsampling convolution on the model in CCFFAM module

Table 3. Influence of downsampling convolution on the model in CCFFAM module

Table 4. Influences of different position detection heads on the model

Table 4. Influences of different position detection heads on the model

Table 5. Influences of different loss functions on the model

Table 5. Influences of different loss functions on the model

Table 6. Comparison of detection results of various small targets on VisDrone2019 dataset

Table 6. Comparison of detection results of various small targets on VisDrone2019 dataset

Table 7. Ablation experiments on VisDrone2019 dataset

Table 7. Ablation experiments on VisDrone2019 dataset

Table 8. Ablation experiments on DOTAv1 dataset

Table 8. Ablation experiments on DOTAv1 dataset

Table 9. Comparison experiments on VisDrone2019 dataset

Table 9. Comparison experiments on VisDrone2019 dataset