Underwater Object Detection Algorithm Based on Improved CenterNet

Net	Stage 1	Stage 2	Stage 3	Stage 4	Resolution
Subnet_1	$[\begin{array}{l} 1 \times 1,64 \\ 3 \times 3,64 \\ 1 \times 1,256 \end{array}] \times 4 \times 1$	$[\begin{array}{l} 3 \times 3,32 \\ 3 \times 3,32 \end{array}] \times 4 \times 1$	$[\begin{array}{l} 3 \times 3,32 \\ 3 \times 3,32 \end{array}] \times 4 \times 4$	$[\begin{array}{l} 3 \times 3,32 \\ 3 \times 3,32 \end{array}] \times 4 \times 3$	4×
Subnet_2		$[\begin{array}{l} 3 \times 3,64 \\ 3 \times 3,64 \end{array}] \times 4 \times 1$	$[\begin{array}{l} 3 \times 3,64 \\ 3 \times 3,64 \end{array}] \times 4 \times 4$	$[\begin{array}{l} 3 \times 3,64 \\ 3 \times 3,64 \end{array}] \times 4 \times 3$	8×
Subnet_3			$[\begin{array}{l} 3 \times 3,128 \\ 3 \times 3,128 \end{array}] \times 4 \times 4$	$[\begin{array}{l} 3 \times 3,128 \\ 3 \times 3,128 \end{array}] \times 4 \times 3$	16×
Subnet_4				$[\begin{array}{l} 3 \times 3,256 \\ 3 \times 3,256 \end{array}] \times 4 \times 3$	32×

Table 2. PASCAL VOC dataset test results

View table

Table 2. PASCAL VOC dataset test results

Algorithm	Bacbone nework	Size	GPU	mAP /%	FPS
Fast R-CNN	VGG-16	~1000×600	Tian X	70.0	0.5
Reference ［21］	ResNet-50		RTX 2080Ti	72.6
Faster R-CNN	ResNet-101	~1000×600	Tian X	76.4	5.0
SSD300	VGG-16	300×300	Tian X	77.1	45
SSD	VGG-16	320×320	Tian X	77.5	11.2
YOLOv3	Darknet-53	554×554	Tian X	79.3	26.0
RetinaNet	ResNet-101	~1000×600	RTX2080 S	75.3	8.8
Proposed algorithm	FA-HRNet	384×384	RTX2080 S	78.1	7.0
Proposed algorithm	FA-HRNet	512×512	RTX2080 S	79.5	6.7

Table 3. Comparison of detection accuracy and speed with CenterNet algorithm
View table
Table 3. Comparison of detection accuracy and speed with CenterNet algorithm
Algorithm Net Scallop /% Holothurian /% Starfish /% Echinus /% mAP /% FPS
CenterNet Hourglass-104 57.4 71.8 86.0 88.5 75.9 5.2
Proposed algorithm FA-HRNet 60.2 73.3 85.6 89.7 77.4 7.0

Table 4. Comparison of model complexity with CenterNet algorithm
View table
Table 4. Comparison of model complexity with CenterNet algorithm
Algorithm Net Ipout size Size /MB Params /MB GFLOPs /10⁹
CenterNet Hourglass-104 384×384 765.7 191.2 164.5
Proposed algorithm FA-HRNet 384×384 123.0 30.4 32.6

Table 5. Performance comparison with mainstream object detection algorithms

View table

Table 5. Performance comparison with mainstream object detection algorithms

Algorithm	Bacbone nework	Size /MB	Params /MB	GFLOPs /10⁹	mAP /%	FPS
Faster R-CNN	ResNet-101+FPN	552.8	59.5	91.8	73.7	3.5
Reference ［22］	ResNet-50				72.6
SSD	VGG-16	92.61	24.2	30.6	68.6	16.0
YOLOv3	Darknet-53	246.5	61.5	32.8	73.3	15.0
RetinaNet	ResNet-101	228.5	55.2	100.6	72.2	8.8
CornerNet	Hourglass-104	804.6	201.0	453.0	49.0	3.1
ExtremeNet	Hourglass-104	794.1	198.3	229.9	53.5	2.3
CenterNet	Hourglass-104	765.7	191.2	164.5	75.9	5.2
Proposed algorithm	FA-HRNet	123.0	30.4	32.6	77.4	7.0

Table 6. Influence of different modules on detection performance
View table
Table 6. Influence of different modules on detection performance
No. HRNet BAM FFM Size /MB Params /MB GFLOPs /10⁹ mAP /% FPS
1 — — — 765.7 191.24 164.53 75.9 5.2
2 √ — — 115.2 28.67 24.16 74.7 8.2
3 √ √ — 115.6 28.74 24.17 76.2 7.9
4 √ √ √ 123.0 30.36 32.55 77.4 7.0

Tools

Get Citation

Copy Citation Text

Rongrong Wang, Zhongyun Jiang. Underwater Object Detection Algorithm Based on Improved CenterNet[J]. Laser & Optoelectronics Progress, 2023, 60(2): 0215001

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites