Lightweight Object Detection Network Based on Convolutional Neural Network

Yequn Cheng; Yan Wang; Yuying Fan; Baoqing Li

doi:10.3788/LOP202158.1610023

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 16 >Page 1610023 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 16, 1610023 (2021)

Lightweight Object Detection Network Based on Convolutional Neural Network

Yequn Cheng^1、2, Yan Wang^1、2, Yuying Fan^1、2, and Baoqing Li^1、*

Author Affiliations

¹Key Laboratory of Microsystem Technology, Shanghai Institute of Microsystem and Information Technology, Chinese Academy of Sciences, Shanghai 201800, China

²University of Chinese Academy of Sciences, Beijing 100049, China

show less

DOI: 10.3788/LOP202158.1610023 Cite this Article Set citation alerts

Yequn Cheng, Yan Wang, Yuying Fan, Baoqing Li. Lightweight Object Detection Network Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2021, 58(16): 1610023 Copy Citation Text

show less

Fig. 1. Depthwise separable convolution

Download full size

Fig. 2. Channel feature interweaving module

Download full size

Fig. 3. Adaptive multi-scale weighted feature fusion module

Download full size

Fig. 4. Spatial pyramid pooling

Download full size

Fig. 5. Overall network structure of BENet

Download full size

Fig. 6. Loss curve of BENet and YOLOv3

Download full size

Fig. 7. Comparison of the detection results of different algorithms on the VOC dataset. (a) YOLOv3; (b) BENet; (c) YOLOv3 tiny

Download full size

Input size	Operator	Stride	N	Output size
416×416×3	Conv2d	2	1	208×208×32
208×208×32	Bottleneck	1	1	208×208×16
208×208×16	Bottleneck	2	2	104×104×24
104×104×24	Feature interweaving	1	2	104×104×24
104×104×24	Bottleneck	2	3	52×52×32
52×52×32	Feature interweaving	1	2	52×52×32
52×52×32	Bottleneck	2	4	26×26×64
26×26×64	Bottleneck	1	3	26×26×96
26×26×96	Feature interweaving	1	2	26×26×96
26×26×96	Bottleneck	2	3	13×13×160
13×13×160	Bottleneck	1	1	13×13×320
13×13×320	Feature interweaving	1	2	13×13×320
13×13×320	Conv2d	1	1	13×13×1280

Table 1. Related parameters of I-MobileNetv2

Algorithm	Input size	Backbone	mAP/%	FPS	Params/10⁶
Faster-RCNN	600×1000	VGG	73.2	7	138.5
SSD	512×512	VGG	76.8	22	33.1
DSSD	513×513	ResNet-101	81.5	5.5
SSDLite	300×300	MobileNet	72.7	56	6.8
R-FCN	600×1000	ResNet-101	80.5	9
RFBNet	512×512	VGG	82.2	38	34.5
YOLO	448×448		66.4	20	86.7
YOLOv3 tiny	544×544		61.1	116	8.5
YOLOv3	544×544	Darknet 53	80.4	26.5	62.3
BENet	544×544	I-MobileNetv2	78.4	49	7.9

Table 2. Comparison of the results of different object detection algorithms on PASCAL VOC dataset

Algorithm	Model size /MB	BFLOPs
YOLOv3 tiny (416)	34	5.56
YOLOv3(416)	236	65.86
BENet(416)	32	6.31

Table 3. Comparison of model size and number of calculations between BENet and YOLOv3

BENet baseline	With MobileNetv2	With I-MobileNetv2	With SPP	With weighted feature fusion	mAP /%
√	√				75.1
√		√			76.3
√		√	√		77.6
√		√	√	√	78.4

Table 4. Ablation experiment on PASCAL VOC dataset

Yequn Cheng, Yan Wang, Yuying Fan, Baoqing Li. Lightweight Object Detection Network Based on Convolutional Neural Network[J]. Laser & Optoelectronics Progress, 2021, 58(16): 1610023

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information