Surface Corrosion Detection of Quayside Crane Based on Improved MobileNetV2SSDLite

Dong Han; Gang Tang; Zhengkun Zhao

doi:10.3788/LOP202158.1615006

Laser & Optoelectronics Progress, Volume. 58, Issue 16, 1615006(2021)

Surface Corrosion Detection of Quayside Crane Based on Improved MobileNetV2SSDLite

Dong Han^1、*, Gang Tang¹, and Zhengkun Zhao²

Author Affiliations

¹School of Logistics Engineering, Shanghai Maritime University, Shanghai 201306, China

²Department of Electronic Engineering, University of York, YO105DD, UK

show less

Abstract Get PDF(in Chinese)

Figures & Tables(17)

Fig. 1. Structure of MobileNetV2SSDLite detection model

Download full size

Fig. 2. Process of standard convolution and depth separable convolution. (a) Standard convolution; (b) depth separable convolution

Download full size

Fig. 3. Different residuals. (a) Residual structure; (b) inverted residual structure

Download full size

Fig. 4. Inverted residual block.(a) Stride is 1; (b) stride is 2

Download full size

Fig. 5. Structure of improved network

Download full size

Fig. 6. Images after data enhancement. (a) Original image; (b) color distortion; (c) random cropping; (d) horizontal flip; (e) random sampling

Download full size

Fig. 7. Performance comparison of different models

Download full size

Fig. 8. Number of network parameters and number of floating point operations

Download full size

Fig. 9. Performance curves of various networks under different conditions. (a) Crack; (b) erosion; (c) overall

Download full size

Fig. 10. Hard samples and easy sample. (a) Difficult sample 1; (b) simple sample; (c) difficult sample 2; (d) difficult sample 3

Download full size

Fig. 11. Detection results of different networks. (a) image 1; (b) image 2; (c) image 3; (d) image 4

Download full size

Fig. 12. Detection results of corrosion of quay bridge. (a) Banded corrosion; (b) pitting corrosion; (c) block corrosion

Download full size

Table 1. Size of network priori box

View table

Table 1. Size of network priori box

Network structure	Layer	Size/(pixel, pixel)
Original SSD	conv 4_3, conv 7, conv 8_2, conv 9_2, conv 10_2, conv 11_2	(30, 60), (60, 111),(111, 162), (162, 213), (213, 264), (264, 315)
MobileNetV2SSDLite	conv 13, conv 1, layer 19_2_2, layer 19_2_3, layer 19_2_4, layer 19_2_5	(60, 60), (105, 150), (150, 195), (195, 240), (240, 285), (285, 300)
Ours-MobileNetV2SSDLite	conv 13, conv 1, layer 19_2_2, layer 19_2_3, layer 19_2_4	(51.2, 51.2), (89.6,128.0), (128.0,166.4), (166.4, 204.8), (204.8, 256.0)

Table 2. Parameters of basic network structure
View table
Table 2. Parameters of basic network structure
Backbone Bottleneck group Expand ratio Times of repetition Total bottleneck
Original MobileNetV2 7 1, 6, 6, 6, 6, 6, 6 1, 2, 3, 4, 3, 3, 1 17
Ours-MobileNetV2 8 1, 6, 6, 6, 6, 6, 6, 6 3, 2, 1, 2, 1, 1, 2, 2 14

Table 3. Parameters of network structure

View table

Table 3. Parameters of network structure

Network structure	Size of input image /(pixel×pixel)	Number of default boxes	Feature map size /(pixel×pixel)	Number of prior boxes
Original SSD	300×300	4, 6, 6, 6, 4, 4	38×38, 19×19, 10×10, 5×5, 3×3, 1×1	8732
MobileNetV2SSDLite	300×300	3, 6, 6, 6, 6, 6	38×38, 19×19, 10×10, 5×5, 3×3, 1×1	7308
Ours-MobileNetV2SSDLite	256×256	3, 6, 6, 6, 6	16×16, 8×8, 4×4, 2×2, 1×1	1278

Table 4. Performance comparison of different networks

View table

Table 4. Performance comparison of different networks

Network	Params /10⁶	FLOPs /10⁹	mAP /%	FPS
ShuffleNet-SSD	8.26	1.23	74.13	42
YOLOV3-Tiny	7.90	0.81	72.14	40
MobileNetV1SSD	5.64	0.79	73.63	40
SqueezeNet-SSD	5.52	0.51	76.30	43
MobileNetV2SSDLite	3.32	0.64
MobileNetV2SSDLiteV1	3.30	0.44	77.40	40
MobileNetV2SSDLiteV2	0.96	0.12	75.62	45

Table 5. Impact of various methods on network performance

View table

Table 5. Impact of various methods on network performance

Data enhancement	Half channel	No bias	Multi-scale	mAP /%	Params /10⁶	FPS
No	No	No	No	73.78	3.30	40
No	No	No	Yes	76.43	3.30	40
Yes	No	No	No	76.35	3.30	40
No	Yes	No	No	72.33	1.58	44
No	No	Yes	No	71.98	2.68	41
Yes	No	No	Yes	77.40	3.30	40
Yes	Yes	Yes	Yes	75.62	0.96	45

Tools

Get Citation

Copy Citation Text

Dong Han, Gang Tang, Zhengkun Zhao. Surface Corrosion Detection of Quayside Crane Based on Improved MobileNetV2SSDLite[J]. Laser & Optoelectronics Progress, 2021, 58(16): 1615006

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites