Multiscale Feature Fusion and Anchor Adaptive Object Detection Algorithm

Table 1. Experimental environment configuration
View table
Table 1. Experimental environment configuration
Hardware Software
CPU：AMDR5-4600H Operating System
GeForce GTX1650Ti-4G Ubuntu 20.0
GPU：NVIDIA RTX1650 Frame：mmdetection 2.0
512 GB SSD Language：python

Table 2. AP of different feature fusion methods for twenty target categories

View table

Table 2. AP of different feature fusion methods for twenty target categories

Category	FPN	PA-FPN	B-FPN	Proposed method	Variation1 /percent	Variation 2 /percent
mAP	70.3	72.745	72.365	72.85	+0.105	+0.485
Aero plane	75.9	79.0	79.0	77.8	-1.2	-1.2
Bicycle	77.2	80.0	79.8	79.7	-0.3	-0.1
Bird	69.1	71.1	70.8	70.2	-0.9	-0.6
Boat	57.9	60.1	59.2	60.9	+0.8	+1.7
Bottle	56.4	58.9	57.8	59.2	+0.3	+1.4
Bus	75.2	82.9	78.5	78.0	-4.9	-0.5
Car	79.3	80.5	80.3	80.7	+0.2	+0.4
Cat	81.7	86.3	83.9	85.2	-1.1	+1.3
Chair	52.2	54.1	54.4	54.3	+0.2	-0.1
Cow	75.8	77.4	81.2	77.3	-0.1	-3.9
Dining table	63.4	65.4	66.7	65.7	+0.3	-1.0
Dog	82.9	85.3	83.6	85.2	-0.1	+1.6
Horse	78.8	80.7	81.1	80.7	0	-0.4
Motorbike	76.2	78.9	79.2	80.2	+1.3	+1.0
Person	77.9	79.3	79.2	79.6	+0.3	+0.4
Potted plant	42.2	44.1	43.2	45.2	+1.1	+2.0
Sheep	71.1	73.4	72.3	75.4	+2.0	+3.1
Sofa	68.2	69.9	69.6	73.0	+3.1	+3.1
Train	77.6	78.8	78.3	79.9	+1.1	+1.6
Tv monitor	67.0	68.8	69.2	68.8	0	-0.4

Table 3. Ablation experiment results
View table
Table 3. Ablation experiment results
Group ResNet50 B-PAFPN Adaptive-anchor mAP /%
G1 × × × 69.9
G2 √ × × 70.3
G3 √ √ × 72.85
G4 √ √ √ 73.5

Table 4. mAP comparison of different algorithms
View table
Table 4. mAP comparison of different algorithms
Algorithm Backbone mAP /%
Faster R-CNN VGG16 69.9
Faster R-CNN^［25］ Res2Net 74.4
Libra R-CNN^［19］ ResNet50 72.6
YOLO^［26］ Dark-Net 63.4
SSD^［27］ VGG16 68.0
Proposed algorithm ResNet50 73.5

Tools

Get Citation

Copy Citation Text

Runmei Zhang, Lijun Bi, Fangbin Wang, Bin Yuan, Gu'an Luo, Huaizhen Jiang. Multiscale Feature Fusion and Anchor Adaptive Object Detection Algorithm[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215019

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites