Application of Deep Convolution Neural Network in Crack Identification

Figures & data

Figure 1. Structure diagram of convolutional neural network.

Figure 2. An example of convolution computation, the step size is 1. The upper part is the convolution operation of multiple pictures, and the lower part is the visualization of one step in the convolution calculation.

Figure 3. An example of Max pooling, the step size is 1.

Figure 4. An example of Avg pooling, the step size is 1.

Figure 5. Sigmoid function image.

Figure 6. ReLU function image.

Figure 7. Neuron (Perceptron).

Figure 8. A simple example of fully connected network.

Figure 9. All above are cracked images. From top to bottom: unclear and hard to detect cracks; small cracks; obvious and clear cracks.

Figure 10. All above are uncracked images. The upper part are smooth uncracked images, the lower part are rough uncracked images.

Figure 11. An image of the original crack in the dataset.

Figure 12. Four images are obtained by image enhancing.

Table 1. Number of images before augmentation

Class type	Train folder	Test folder	Total(class)
Un-cracked	1620	405	2025
Cracked	9276	2319	11,595
Total(folder)	10,896	2724	13,620

Table 2. Number of images after enhancement

Class type	Train folder	Test folder	Total(class)
Un-cracked	37,932	9676	47,608
Cracked	40,773	10,131	50,904
Total(folder)	78,705	19,807	98,512

Table

Model	Total layers	Trainable params	Total params
VGG19	29	20,189,761	20,190,785
InceptionV3	318	22,330,017	22,368,545
ResNet50	182	24,096,257	24,153,473
Xception	139	21,368,617	21,427,241
CNN	16	500,897	501,153
FCNN	13	2,111,425	2,111,489

Table 4. Loss, RMSE and accuracy of cross-validation results for each prediction model

Model	Loss	RMSE	Accuracy
VGG19	0.1790	0.2700	0.9297
InceptionV3	0.0958	0.2081	0.9639
ResNet50	0.1310	0.2616	0.9479
Xception	0.1297	0.2213	0.9520
CNN	0.4251	0.4283	0.7955
FCNN	0.6375	0.4776	0.6394

Figure 13. Change trend of loss.

Figure 14. Change trend of RMSE.

Figure 15. Change trend of accuracy.

Table 5. Results of precision, recall and F₁-score

Model	P	R	F1
Xception	0.9650	0.9678	0.9664
InceptionV3	0.9131	0.9402	0.9264
ResNet50	0.9004	0.9399	0.9197
VGG19	0.8829	0.9146	0.8985

Figure 16. Confusion matrix results of four models.

Table 6. Precision, recall and F₁-score of cross-datasets research

Dataset	P	R	F1
Dataset-1	0.9823	0.9918	0.9870
Dataset-2	0.9904	0.9869	0.9886

Figure 17. Confusion matrix results of cross-datasets research.