Lightweight YOLOv3 Algorithm for Small Object Detection

Fig. 12. Comparison of channel numbers before and after channel pruning. (a) Channel number comparison of backbone before and after channel pruning; (b) channel number comparison before and after channel pruning besides backbone

Download full size

Fig. 13. Layer pruning result graph

Download full size

Table 0. [in Chinese]

View table

Table 0. [in Chinese]

Algorithm 1 Adaptive adjustment of sparsity factor： $λ$
set：the group number of parameter $γ$ is K，the sparsity factor is $λ$ ，the number of iterations is epochs Begin
While epoch < epochs do
While k < K do
If 80% of a parameter in Group K is 0
The sparsity factor of this group is $λ_{k} = λ \times 0.01$
End
End
End
End

Table 1. Channel pruning results

View table

Table 1. Channel pruning results

Name	Size /MB	Number of parameters /MB	GFLOPS	mAP@0.5	F1	FPS
YOLOv3-CS	215.5	51.2	65.6	0.903	0.899	78
YOLOv3-CS（sparse train）	215.5	51.2	65.6	0.906	0.876	78
YOLOv3-CS-CP（pruning rate）	10.8	2.68	14	0.896	0.873	147
YOLOv3-CS-CP（fine-tuned after layer pruned）	10.8	2.68	14	0.902	0.875	147

Table 2. Layer pruning results
View table
Table 2. Layer pruning results
Name Size /MB Number of parameters /MB GFLOPS mAP@0.5 F1 FPS
YOLOv3-CS-LP（layer pruned） 115.7 28.91 40.7 0.649 0.575 108
YOLOv3-CS-LP（fine-tuned after layer pruned） 115.7 28.91 40.7 0.902 0.855 108

Table 3. Comparison of proposed pruning method, literature [25], and literature [24] pruning methods on YOLOv3-CS

View table

Table 3. Comparison of proposed pruning method, literature [25], and literature [24] pruning methods on YOLOv3-CS

Name	Size /MB	Number of parameters /MB	GFLOPS	mAP@0.5	F1	FPS
YOLOv3-CS	215.5	51.23	65.6	0.903	0.899	78
YOLOv3-CS（70% pruning rate）^［25］	12.7	2.97	8.7	0.896	0.875	111
YOLOv3-CS（90% pruning rate）^［24］	28.3	6.67	8.2	0.899	0.885	122
YOLOv3-CSP	8.8	2.18	11.6	0.901	0.889	213

Table 4. Comparison of proposed pruning method, literature [25], and literature [24] pruning methods on YOLOv3

View table

Table 4. Comparison of proposed pruning method, literature [25], and literature [24] pruning methods on YOLOv3

Name	Size /MB	Number of parameters /MB	GFLOPS	mAP@0.5	F1	FPS
YOLOv3	246.5	61.63	65.7	0.848	0.825	85
YOLOv3（70% pruning rate）^［25］	14.5	3.58	8.8	0.834	0.789	121
YOLOv3（90% pruning rate）^［24］	31.2	8.01	8.3	0.842	0.819	141
YOLOv3（our）	10.1	2.62	11.6	0.846	0.824	224

Table 5. Lightweight model comparison

View table

Table 5. Lightweight model comparison

Name	Size /MB	Number of parameters /MB	GFLOPS	mAP@0.5	F1	FPS
MobileNet-YOLOv3	95.6	23.82	14.5	0.867	0.836	116
GhostNet-YOLOv3	94.2	23.51	14.1	0.861	0.828	133
SlimYOLOv3-SPP3-90^［24］	32.4	8.02	9.97	0.892	0.871	133
YOLOv3-tiny	33.3	8.67	5.5	0.843	0.858	322
YOLOv4-tiny	24.4	6.07	13.2	0.869	0.861	243
YOLO-fastest	1.4	0.29	0.8	0.687	0.660	313
YOLO-fastest-xl	3.6	0.84	2.3	0.720	0.637	204
YOLOv3-CSP	8.8	2.18	11.6	0.901	0.889	213

Tools

Get Citation

Copy Citation Text

Guanrong Zhang, Xiang Chen, Yu Zhao, Jianjun Wang, Guobiao Yi. Lightweight YOLOv3 Algorithm for Small Object Detection[J]. Laser & Optoelectronics Progress, 2022, 59(16): 1610008

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites