Bridge Crack Detection Algorithm Based on Image Processing under Complex Background

Fig. 1. Schematic of dataset amplification of bridge crack images. (a) Original image; (b) horizontal flip; (c) vertical flip; (d) linear transformation; (e) spatial filtering transformation

Download full size

Fig. 2. Generative model

Download full size

Fig. 3. Discriminant model

Download full size

Fig. 4. Schematic of 4-layer DenseBlock

Download full size

Fig. 5. Schematic of detection of high-resolution image

Download full size

Fig. 6. Visualization comparison of cracks generated by DCGAN and BCIGM. (a) Nepoch=01; (b) Nepoch=03; (c) Nepoch=16; (d) Nepoch=25

Download full size

Fig. 7. Visualization comparison of cracks generated by ReLU and SeLU. (a) Nepoch=01; (b) Nepoch=03; (c) Nepoch=16; (d) Nepoch=25

Download full size

Fig. 8. Visualization comparison of experimental results with and without dataset amplification. (a) Original image; (b)label; (c) without dataset amplification; (d) with dataset amplification

Download full size

Fig. 9. Comparison of crack detection results between existing algorithms and proposed algorithm. (a) Original image; (b) label; (c) threshold segmentation algorithm; (d) Canny algorithm; (e) NB-CNN algorithm; (f) random structure forest algorithm; (g) proposed algorithm

Download full size

Fig. 10. Partial crack detection results by proposed algorithm. (a) Scene 1; (b) scene 2; (c) scene 3

Download full size

Table 1. Network structure parameters of BCISM

View table

Table 1. Network structure parameters of BCISM

Name oflayer	Size ofkernel /(pixel×pixel)	Stride /pixel	Size of output featuremap /(pixel×pixel)	Number offeature map
Inputlayer	-	-	256×256	3
Convolution	Convolution 5×5	1	256×256	48
DenseBlock	Convolution [3×3]×4	1	256×256	96
Transition Down	Convolution 1×1	1	256×256	96
Transition Down	Max pooling 2×2	2	128×128	96
DenseBlock	Convolution [3×3]×5	1	128×128	156
Transition Down	Convolution 1×1	1	128×128	156
Transition Down	Max pooling 2×2	2	64×64	156
DenseBlock	Convolution [3×3]×7	1	64×64	240
Transition Down	Convolution 1×1	1	64×64	240
Transition Down	Max pooling 2×2	2	32×32	240
DenseBlock	Convolution [3×3]×10	1	32×32	360
Transition Down	Convolution 1×1	1	32×32	360
Transition Down	Max pooling 2×2	2	216×16	360
DenseBlock	Convolution [3×3]×12	1	16×16	504
Transition Up	Deconvolution 3×3	2	32×32	504
DenseBlock	Convolution [3×3]×10	1	32×32	624
Transition Up	Deconvolution 3×3	2	64×64	624
DenseBlock	Convolution [3×3]×7	1	64×64	444
Transition Up	Deconvolution 3×3	2	128×128	444
DenseBlock	Convolution [3×3]×5	1	128×128	300
Transition Up	Deconvolution 3×3	2	256×256	300
DenseBlock	Convolution [3×3]×4	1	256×256	204
Convolution	Convolution 1×1	1	256×256	2
Softmax	-	-	-	-

Table 2. Proportion of number of different types of images in total dataset
View table
Table 2. Proportion of number of different types of images in total dataset
Type ofpicture Simplebackground Backgroundwith obstacles Backgroundwith largearea of stains
Number ofpictures 10449 13275 5710
Proportion /% 35.5 45.1 19.4

Table 3. Influence of BCIGM on training speed under different conditions
View table
Table 3. Influence of BCIGM on training speed under different conditions
Condition Time /s
Without 1×1 convolution kernel 5.4801
With 1×1 convolution kernel 5.4312

Table 4. Effect of dataset amplification on experimental results
View table
Table 4. Effect of dataset amplification on experimental results
Number oftraining samples With or withoutdataset amplification Number ofverification samples P_Precision /% P_Recall /%
1183 Without dataset amplification 156 13.5 17.9
29434 With dataset amplification 156 92.9 92.6

Table 5. Comparison of exiting semantic segmentation models and BCISM

View table

Table 5. Comparison of exiting semantic segmentation models and BCISM

Model	Pre-training	Parameter /M	P_Precision /%	P_Recall /%	P_{F1_Score} /%	Time /s
SegNet	True	29.5	74.0	78.5	76.2	0.5823
FCN8	True	134.5	86.9	83.4	85.1	0.3739
DeepLab	True	37.3	82.6	80.9	81.7	0.9751
FC-DenseNet56 (k=12)	False	1.5	89.8	87.6	88.7	0.1685
FC-DenseNet67 (k=16)	False	3.5	89.0	88.8	88.9	0.2635
FC-DenseNet103 (k=16)	False	9.4	93.0	92.1	92.5	0.2795
BCISM (k=12)	False	2.8	92.9	92.6	92.8	0.1998

Tools

Get Citation

Copy Citation Text

Liangfu Li, Ruiyun Sun. Bridge Crack Detection Algorithm Based on Image Processing under Complex Background[J]. Laser & Optoelectronics Progress, 2019, 56(6): 061002

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites