Research on Train Key Components Detection Based on Improved RetinaNet

Table 1. Computing environment configuration
View table
Table 1. Computing environment configuration
Name Configuration
CPU Intel i5-8300H
GPU GTX1060，6GB
System Windows10
Cuda/cuDNN 9.0/9.0
Framework Python3/Tensorflow1.4

Table 2. Validation results on PASCAL VOC dataset (AP)

View table

Table 2. Validation results on PASCAL VOC dataset (AP)

Method	mAP	Aero	Bike	Bird	Boat	Bottle	Bus	Car	Cat	Chair	Cow
RetinaNet	0.770	0.88	0.84	0.80	0.67	0.50	0.81	0.85	0.93	0.56	0.83
Proposed	0.776	0.89	0.84	0.81	0.66	0.52	0.81	0.86	0.94	0.57	0.85
Method	Table	Dog	Horse	Mbike	Person	Plant	Sheep	Sofa	Train	TV
RetinaNet	0.64	0.90	0.83	0.84	0.81	0.52	0.77	0.74	0.89	0.79
Proposed	0.66	0.90	0.85	0.83	0.82	0.52	0.77	0.73	0.89	0.79

Table 3. Key component dataset
View table
Table 3. Key component dataset
Category Training set Testing set
Objects m-s ratio /% Objects m-s ratio /%
Brake 574 91 233 92
U-lock 194 94 99 89
Cotter pin 235 88 104 96
Bolt 4542 73 1436 64
Hexagonal lock 164 98 87 97
Nameplate 339 63 130 81

Table 4. Influence of different feature layers on detection effect
View table
Table 4. Influence of different feature layers on detection effect
P3、P4、P5、P6、P7 P3、P4、P5 mAP Weight /MB
√ 0.952 116
√ 0.951 109

Table 5. AP comparison between proposed method and RetinaNet
View table
Table 5. AP comparison between proposed method and RetinaNet
Model mAP Brake U-lock Cotter pin Bolt Hexagonal lock Nameplate
RetinaNet 0.951 0.94 0.98 0.93 0.95 0.97 0.94
RetinaNet+RFB 0.958 0.95 0.99 0.95 0.95 0.96 0.95
Proposed 0.964 0.95 0.99 0.96 0.95 0.97 0.96

Table 6. Comparison of detection speed and efficiency of different models
View table
Table 6. Comparison of detection speed and efficiency of different models
Method mAP Inference time /s
YOLOV3 0.85 0.08
YOLOV4 0.93 0.10
FasterRCNN 0.92 0.25
FCOS 0.89 0.10
SSD 0.93 0.10
RetinaNet 0.95 0.11
Proposed 0.96 0.12

Tools

Get Citation

Copy Citation Text

Kai Yang, Rui Li, Lin Luo, Liming Xie. Research on Train Key Components Detection Based on Improved RetinaNet[J]. Laser & Optoelectronics Progress, 2022, 59(12): 1215006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites