Deep Multi-Scale Feature Fusion Target Detection Algorithm Based on Deep Learning

Xin Liu; Siyi Chen; Xiaolong Chen; Xinhao Du

doi:10.3788/LOP202158.1210029

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 12 >Page 1210029 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 12, 1210029 (2021)

Deep Multi-Scale Feature Fusion Target Detection Algorithm Based on Deep Learning

Xin Liu, Siyi Chen^***, Xiaolong Chen^**, and Xinhao Du^*

Author Affiliations

School of Automation and Electronic Information, Xiangtan University, Xiangtan, Hunan 411105, China

show less

DOI: 10.3788/LOP202158.1210029 Cite this Article Set citation alerts

Xin Liu, Siyi Chen, Xiaolong Chen, Xinhao Du. Deep Multi-Scale Feature Fusion Target Detection Algorithm Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210029 Copy Citation Text

show less

Fig. 1. Algorithm framework of SSD

Download full size

Fig. 2. Algorithm framework of DMSFFD

Download full size

Fig. 3. Feature fusion structure. (a) First feature fusion; (b) second feature fusion

Download full size

Fig. 4. Test result of occluded object by each algorithm. (a) SSD;(b) DMSFFD

Download full size

Fig. 5. Positioning result of target by each algorithm. (a) SSD;(b) DMSFFD

Download full size

Fig. 6. Detection result of small target by each algorithm. (a) SSD;(b) DMSFFD

Download full size

Fig. 7. Detection result of multiple occluded objects by each algorithm. (a) SSD;(b) DMSFFD

Download full size

SSD		DMSFFD		Size /( pixel×pixel)
Feature map	Dimention	Feature map	Dimention	Size /( pixel×pixel)
Conv4_3	512	F1_conv	256	38×38
Conv7	1024	F2_conv	256	19×19
Conv8_2	512	F3_conv	256	10×10
Conv9_2	256	F4_conv	256	5×5
Conv10_2	256	F5_conv	256	3×3
Conv11_2	256	F6_conv	256	1×1

Table 1. Feature map details

Image	DSSD	SSD	Faster	DMSFFD
Aero	89.9	88.4	76.5	90.7
Bike	87.9	86.0	79.0	89.7
Bird	85.5	78.9	70.9	90.3
Boat	78.4	75.8	65.5	88.0
Bottle	53.9	48.8	52.1	70.3
Bus	88.6	86.8	83.1	90.7
Car	86.2	84.1	84.7	90.0
Cat	91.9	90.9	86.4	90.9
Chair	71.1	69.1	52.0	87.1
Cow	89.5	88.0	81.9	90.9
Table	78.7	78.4	65.7	89.0
Dog	91.3	90.5	84.8	90.8
Horse	89.6	89.0	84.6	90.6
Motor	88.4	86.8	77.5	90.6
Person	79.2	76.2	76.7	84.5
Plant	61.8	57.0	38.8	81.7
Sheep	78.0	72.7	73.6	82.7
Sofa	89.9	88.3	73.9	93.9
Train	93.2	92.0	83.0	97.0
TV	84.4	83.4	72.6	90.6

Table 2. Test results on VOC2007 dataset unit:%

Algorithm	DSSD	SSD	Faster	DMSFFD
mAP/%	81.4	80.5	73.2	88.5

Table 3. mAP comparison of algorithms on VOC2007 test set

Algorithm	DSSD	SSD	DMSFFD
Detection time	9.5	63	38

Table 4. Detection speed comparison frame/s

Image	DSSD	SSD	Faster	DMSFFD
Aero	87.3	87.0	84.9	90.7
Bike	84.3	83.8	79.8	90.3
Bird	79.4	78.8	74.3	90.0
Boat	69.6	68.0	53.9	84.0
Bottle	56.8	55.4	49.8	71.9
Bus	86.7	84.0	77.5	90.6
Car	76.5	75.0	75.9	84.1
Cat	92.9	90.8	88.5	90.9
Chair	69.5	65.0	45.6	83.9
Cow	81.3	79.7	77.1	90.0
Table	74.3	72.6	55.3	85.1
Dog	91.5	90.3	86.9	90.9
Horse	88.6	88.2	81.7	90.7
Motor	88.6	86.8	80.9	90.5
Person	82.1	79.5	79.6	86.3
Plant	60.3	59.4	40.1	78.5
Sheep	79.6	77.8	72.6	87.3
Sofa	79.7	79.5	60.9	90.1
Train	88.2	88.1	81.2	90.8
TV	79.9	78.8	61.5	90.6

Table 5. Test results on VOC2012 test set %

Algorithm	DSSD	SSD	Faster	DMSFFD
mAP	79.9	78.4	70.4	87.4

Table 6. mAP comparison of algorithms on VOC2012 test set unit:%

Xin Liu, Siyi Chen, Xiaolong Chen, Xinhao Du. Deep Multi-Scale Feature Fusion Target Detection Algorithm Based on Deep Learning[J]. Laser & Optoelectronics Progress, 2021, 58(12): 1210029

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information