Lightweight Target Detection Algorithm Based on Adaptive Spatial Feature Fusion

Yujie Luo; Jian Zhang; Liang Chen; Lü Zhang; Wanqing Ouyang; Daiqin Huang; Yuyi Yang

doi:10.3788/LOP202259.0415004

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 4 >Page 0415004 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 4, 0415004 (2022)

Lightweight Target Detection Algorithm Based on Adaptive Spatial Feature Fusion

Yujie Luo, Jian Zhang^*, Liang Chen, Lü Zhang, Wanqing Ouyang, Daiqin Huang, and Yuyi Yang

Author Affiliations

School of Information and Electrical Engineering, Hunan University of Science and Technology, Xiangtan , Hunan 411100, China

show less

DOI: 10.3788/LOP202259.0415004 Cite this Article Set citation alerts

Yujie Luo, Jian Zhang, Liang Chen, Lü Zhang, Wanqing Ouyang, Daiqin Huang, Yuyi Yang. Lightweight Target Detection Algorithm Based on Adaptive Spatial Feature Fusion[J]. Laser & Optoelectronics Progress, 2022, 59(4): 0415004 Copy Citation Text

show less

Fig. 1. Standard convolution and depth separable convolution. (a) Standard convolution; (b) deep convolution; (c) point convolution

Download full size

Fig. 2. PANet structure

Download full size

Fig. 3. Top-down integration

Download full size

Fig. 4. Improved PANet structure

Download full size

Fig. 5. Network output of YOLOv4

Download full size

Fig. 6. New network output

Download full size

Fig. 7. Structure of propoesd network

Download full size

Fig. 8. Comparison of training effect between proposed algorithm and YOLOv4 under a small amount of data samples

Download full size

Fig. 9. Comparison of training effect between proposed algorithm and YOLOv4 under a large number of data samples

Download full size

Fig. 10. Comparison of detection effect between proposed algorithm and YOLOv4.(a)(c)(e)(g) YOLOv4 algorithm;(b)(d)(f)(h) proposed algorithm

Download full size

	Type	Number of filters	Size	Output
	Conv	32	3×3	416×416
	Conv dw	32	3×3/2	208×208
	Conv	64	1×1	208×208
	Conv dw	64	3×3/2	104×104
	Conv	128	1×1	104×104
	Conv dw	128	3×3	104×104
	Conv	128	1×1	104×104
	Conv dw	128	3×3/2	52×52
	Conv	256	1×1	52×52
	Conv dw	256	3×3	52×52
	Conv	256	1×1	52×52
	Conv dw	256	3×3/2	26×26
	Conv	512	1×1	26×26
5×	Conv dw	512	3×3	26×26
5×	Conv	512	1×1	26×26
	Conv dw	512	3×3/2	13×13
	Conv	1024	1×1	13×13
	Conv dw	1024	3×3	13×13
	Conv	1024	1×1	13×13

Table 1. MobileNet structure

	Type	Number of filters	Size	Output
	Convolutional	32	3×3	416×416
	Convolutional	64	3×3/2	208×208
1×	Convolutional	32	1×1
	Convolutional	64	3×3
	Residual			208×208
	Convolutional	128	3×3/2	104×104
2×	Convolutional	64	1×1
	Convolutional	128	3×3
	Residual			104×104
	Convolutional	256	3×3/2	52×52
8×	Convolutional	128	1×1
	Convolutional	256	3×3
	Residual			52×52
	Convolutional	512	3×3/2	26×26
8×	Convolutional	256	1×1
	Convolutional	512	3×3
	Residual			26×26
	Convolutional	1024	3×3/2	13×13
4×	Convolutional	512	1×1
	Convolutional	1024	3×3
	Residual			13×13

Table 2. CSPDarkNet-53　structure

Algorithm	AP /%			mAP /%	Detection speed /（frame·s^-1）
Algorithm	Mask	No_mask	No_mask_well	mAP /%	Detection speed /（frame·s^-1）
Faster-RCNN	97.77	97.56	95.33	96.88	2
RetinaFace	74.51	97.35	70.45	80.77	6
Attention-RetinaFace	75.76	98.67	73.89	82.77	8
SSD	77.33	75.55	73.34	74.07	9
YOLOv3	81.62	80.96	77.65	80.07	11
YOLOv3-tiny	78.67	77.92	75.33	77.30	15
YOLOv4	88.92	87.34	85.84	87.36	13
Improved_YOLOv4	96.38	96.96	94.44	95.92	19

Table 3. Performance comparison of different algorithms

Grouping	MobileNet	ASFF	Modify_loss	AP /%			mAP /%	Detection speed /（frame·s^-1）
Grouping	MobileNet	ASFF	Modify_loss	Mask	No_mask	No_mask_well	mAP /%	Detection speed /（frame·s^-1）
G1	$\times$	$\times$	$\times$	88.92	87.34	85.84	87.36	13
G2	√	$\times$	$\times$	86.75	86.02	83.98	85.58	21
G3	√	√	$\times$	91.88	92.32	89.45	91.21	20
G4	√	√	√	96.38	96.96	94.44	95.92	19

Table 4. Comparison of ablation experiments

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information