Crack Detection Method of Holistically-Nested Network Based on Feature Enhancement

Shengjun Xu; Ming Hao; Yuebo Meng; Guanghui Liu; Jiuqiang Han

doi:10.3788/LOP202259.1010003

Journals >Laser & Optoelectronics Progress >Volume 59 >Issue 10 >Page 1010003 > Article

Laser & Optoelectronics Progress
Vol. 59, Issue 10, 1010003 (2022)

Crack Detection Method of Holistically-Nested Network Based on Feature Enhancement

Shengjun Xu^1、2, Ming Hao^1、*, Yuebo Meng^1、2, Guanghui Liu¹, and Jiuqiang Han^1、2

Author Affiliations

¹School of Information and Control Engineering, Xi’an University of Architecture and Technology, Xi’an 710055, Shaanxi , China

²Guangdong Artificial Intelligence and Digital Economy Laboratory (Guangzhou), Guangzhou 510320, Guangdong , China

show less

DOI: 10.3788/LOP202259.1010003 Cite this Article Set citation alerts

Shengjun Xu, Ming Hao, Yuebo Meng, Guanghui Liu, Jiuqiang Han. Crack Detection Method of Holistically-Nested Network Based on Feature Enhancement[J]. Laser & Optoelectronics Progress, 2022, 59(10): 1010003 Copy Citation Text

show less

Fig. 1. HEDNet. (a) Structure of HEDNet; (b) structure of deep supervision module

Download full size

Fig. 2. Structure of CFEHNet model

Download full size

Fig. 3. Convolution-deconvolution feature fusion module

Download full size

Fig. 4. Structure of boundary refinement module

Download full size

Fig. 5. Partial dataset after data enhancement

Download full size

Fig. 6. Change of training loss value and verification accuracy rate

Download full size

Fig. 7. Visual comparison between VGG16 and feature fusion module

Download full size

Fig. 8. Visual comparison of effectiveness of deep supervision module

Download full size

Fig. 9. Comparison of partial test results of Bridge_Crack_Image_Data

Download full size

Fig. 10. Comparison of partial test results of CFD

Download full size

Deconvolution stage	Convolution	Channel	Pool	Output
conv_5_3		512	Max，2 $\times$ 2	2M $\times$ 2M
Fuse with conv_4_3	1 $\times$ 1	512		2M $\times$ 2M
D4（4_1，4_2，4_3）	3 $\times$ 3	256	Max，2 $\times$ 2	4M $\times$ 4M
Fuse with conv_3_3	1 $\times$ 1	256		4M $\times$ 4M
D3（3_1，3_2，3_3）	3 $\times$ 3	128	Max，2 $\times$ 2	8M $\times$ 8M
Fuse with conv_2_2	1 $\times$ 1	128		8M $\times$ 8M
D2（2_1，2_2）	3 $\times$ 3	64	Max，2 $\times$ 2	16M $\times$ 16M
Fuse with conv_1_2	1 $\times$ 1	64		16M $\times$ 16M
D1（1_1，1_2）	3 $\times$ 3	64		16M $\times$ 16M

Table 1. Deconvolution network structure and feature fusion parameters

Sample	Predicted positive sample （positive）	Prediction negative sample （negative）
Actually positive sample （positive）	TP	FN
Actually negative sample （negative）	FP	TN

Table 2. Confusion matrix

Condition	Bridge_Crack_Image_Data			Crack Forest Dataset
Condition	Precision	Recall	F₁	Precision	Recall	F₁
CFEHNet without stage 1	0.8534	0.6907	0.7634	0.8615	0.8648	0.8631
CFEHNet without stage 2	0.8756	0.6913	0.7726	0.8805	0.8845	0.8825
CFEHNet without stage 3	0.8942	0.6945	0.7817	0.9058	0.9056	0.9057
CFEHNet without stage 4	0.9134	0.7011	0.7932	0.9104	0.9013	0.9060
CFEHNet	0.9156	0.7063	0.7974	0.9107	0.9022	0.9064

Table 3. Contribution of side network output to result in each stage

Model	Boundary accuracy	Center accuracy	Overall accuracy
CFEHNet without BR	0.7138	0.9412	0.9167
CFEHNet	0.7331	0.9414	0.9214

Table 4. Comparison experiment of boundary thinning module

Algorithm	Precision	Recall	$F_{1}$
HED^［21］	0.9167	0.6503	0.7608
UNet^［23］	0.8945	0.6822	0.7740
SegNet^［24］	0.8745	0.6887	0.7705
CFEHNet without BR	0.9156	0.7063	0.7974
CFEHNet	0.9214	0.7101	0.8021

Table 5. Comparison of quantitative analysis of Bridge_Crack_Image_Data

Algorithm	Precision	Recall	F1-score
HED^［21］	0.8656	0.8699	0.8677
UNet^［23］	0.8811	0.8897	0.8854
SegNet^［24］	0.8816	0.8789	0.8802
CFEHNet without BR	0.9107	0.9022	0.9064
CFEHNet	0.9164	0.9062	0.9113

Table 6. Comparison of quantitative analysis of CFD

Download Citation

Set citation alerts for the article

Tools

Set citation alerts for the article

Save the article for my favorites

Paper Information