Research on Small Target Detection Technology Based on the MPH-SSD Algorithm

Table 1

Performance comparison of the MPH-SSD algorithm and other algorithms on the PASCAL VOC dataset.

Methods	Backbone	Size	mAP (%)	Fps
R-CNN [25]	AlexNet	227 × 227	53.3	—
Fast R-CNN [27]	VGG16	224 × 224	70.0	—
Faster R-CNN [28]	VGG16	1000 × 600	73.2	7.0
Faster R-CNN [28]	ResNet-101	1000 × 600	76.4	2.4
YOLOv1 [30]	GoogleNet	448 × 448	63.4	45.0
SSD300 [31]	VGG16	300 × 300	77.2	46.0
SSD512 [31]	VGG16	512 × 512	78.5	19.0
DSSD321 [32]	ResNet-101	321 × 321	78.6	9.5
DSSD513 [32]	ResNet-101	513 × 513	81.5	5.5
MDSSD300 [33]	VGG16	300 × 300	78.6	32.2
MDSSD512 [33]	VGG16	512 × 512	81.0	14.5
DF-SSD [34]	DenseNet-S-32-1	300 × 300	78.9	11.6
SEFN300 [35]	VGG16	300 × 300	79.6	55.0
SEFN512 [35]	VGG16	512 × 512	81.2	30.0
YOLOv2 [40]	Darknet19	352 × 352	73.7	81.0
YOLOv3 [41]	Darknet19	320 × 320	79.6	52.2
RSSD300 [42]	VGG16	300 × 300	78.5	35.0
RSSD512 [42]	VGG16	512 × 512	80.8	16.6
FSSD300 [43]	VGG16	300 × 300	78.8	65.8
FSSD512 [43]	VGG16	512 × 512	80.9	35.7
RFB300 [44]	VGG16	300 × 300	80.5	83.0
RFB512 [44]	VGG16	512 × 512	82.2	38.0
MPH-SSD300	VGG16	300 × 300	82.1	53.5
MPH-SSD512	VGG16	512 × 512	87.7	24.6