Real-Time Target Detection of Underwater Relics Based on Multigranularity Pruning

Youbo Zhang; Wei Guo; Yue Zhou; Gaofei Xu; Guangwei Li; Hongming Sun

doi:10.3788/LOP202158.1410019

Journals >Laser & Optoelectronics Progress >Volume 58 >Issue 14 >Page 1410019 > Article

Laser & Optoelectronics Progress
Vol. 58, Issue 14, 1410019 (2021)

Real-Time Target Detection of Underwater Relics Based on Multigranularity Pruning

Youbo Zhang^1、2, Wei Guo^2、3、*, Yue Zhou¹, Gaofei Xu², Guangwei Li², and Hongming Sun^2、3

Author Affiliations

¹College of Engineering Science and Technology, Shanghai Ocean University, Shanghai 201306, China

²Institute of Deep-Sea Science and Engineering, Chinese Academy of Sciences, Sanya, Hainan 572000, China

³University of Chinese Academy of Sciences, Beijing 100049, China

show less

DOI: 10.3788/LOP202158.1410019 Cite this Article Set citation alerts

Youbo Zhang, Wei Guo, Yue Zhou, Gaofei Xu, Guangwei Li, Hongming Sun. Real-Time Target Detection of Underwater Relics Based on Multigranularity Pruning[J]. Laser & Optoelectronics Progress, 2021, 58(14): 1410019 Copy Citation Text

show less

YOLOV4 analysis. (a) Three kinds of dimensional feature diagrams and their corresponding anchor schematic diagrams; (b) cluster analysis result

Fig. 1. YOLOV4 analysis. (a) Three kinds of dimensional feature diagrams and their corresponding anchor schematic diagrams; (b) cluster analysis result

Download full size

Images in the dataset. (a) Image collected by UUV; (b) image of “xiaobaijiao” relic; (c) data mining image; (d) “china” category; (e) “chinavase” category; (f) Mosaic online data augmentation

Fig. 2. Images in the dataset. (a) Image collected by UUV; (b) image of “xiaobaijiao” relic; (c) data mining image; (d) “china” category; (e) “chinavase” category; (f) Mosaic online data augmentation

Download full size

Fig. 3. Key frame image selection algorithm

Download full size

Fig. 4. Enhanced results. (a) Original images; (b) CLAHE enhancement; (c) HE enhancement; (d) UCM enhancement; (e) UDCP enhancement

Download full size

Fig. 5. Structure of Comp_YOLOV4

Download full size

Fig. 6. Process of detection

Download full size

Fig. 7. Target detection results of the models

Download full size

Method	Principle	Advantage	Disadvantage
Pruning	Prune unimportant parameters and connections	Flexible operation and small loss of precision	Need fine-tuning, training time increase
Quantization	Reduce parameter bit width	Low computation cost	Accuracy drops too severely to restore
Knowledge distillation	Transfer knowledge from complex models to small models	A small number of parameters are used for calculation	Knowledge transfer standards are difficult to determine and only apply to classification problems
Network architecture search	Automatically search for a network model that meets the parameters and accuracy requirements	High accuracy	High training costs, search indicators are difficult to set

Table 1. Summary of neural network compression methods

Model	N_TP	N_FP	N_FN	P /%	R /%	R_IoU /%	F₁ /%	m_AP /%	L_loss	V /MB
YOLOV3	829	83	321	91	72	79.92	80	77.51	0.4514	246.3
YOLOV3-SPP	652	31	498	95	57	82.27	71	68.34	0.4270	250.5
YOLOV3-tiny	356	319	794	53	31	35.99	39	26.65	1.0500	34.7
YOLOV4	1292	68	210	95	86	86.74	90	88.34	0.4047	256.0

Table 2. Statistics of the basic training performance of models

Model	α	τ_c	τ_lay	ϕ	F₁ /%	m_AP /%	L_loss	V /MB
YOLOV4-m-8000	10^-3	0.10	0.5	0.05	78	75.50	0.7100	18.5
YOLOV4-pr-8000	10^-3	0.10	0.5	0.05	77	77.90	0.4418	18.5
YOLOV4-m-8000	10^-3/10^-4	0.10	0.5	0.05	85	77.98	0.6123	14.3
YOLOV4-pr-8000	10^-3/10^-4	0.10	0.5	0.05	86	80.47	0.5606	14.3
YOLOV4-m-8000	10^-3/10^-4	0.08	0.5	0.1	73	71.32	1.0513	11.3
YOLOV4-pr-8000	10^-3/10^-4	0.08	0.5	0.1	74	73.82	1.1832	11.3