Design and research on pavement crack segmentation based on convolutional neural network

数据集名称	数量	尺度/像素	数据集特点描述
ESAR^[21]	15	768×512	自然状态下动态拍摄的沥青路面；包含阴影干扰问题
AIGLE_RN^[21]	38	991×462	动态拍摄的沥青路面；存在严重的光照不均和背景椒盐噪点问题
		311×462
DeepCrack^[22]	300	544 ×384	混泥土和沥青路面；包含完好路面、受损路面、脏乱路面等场景
GAPs384^[23]	509	640×540	德国沥青路面；包含坑洼，斑块等干扰问题
Crack500^[24]	1 896	640×340	美国天普大学主校区的混泥土和沥青路面；包含车道线、坑洼破损、椒盐噪声等强干扰问题
CFD^[25]	118	480×320	中国北京的沥青路面状况；包含阴影、车道线、油斑和水渍等干扰问题

Table 2. Comparison of loss functions with different weights

View table

View in Article

Table 2. Comparison of loss functions with different weights

损失函数( $w_{2} = 0.1$ )	aAcc	mAcc	F1	P	R
BCE( $w_{1} = 0.2$ )	0.9838	0.9060	0.7704	0.7243	0.8228
BCE( $w_{1} = 0.3$ )	0.9811	0.9257	0.7519	0.6642	0.8663
BCE( $w_{1} = 0.4$ )	0.9790	0.9350	0.7440	0.6402	0.8881
BCE( $w_{1} = 0.5$ )	0.9773	0.9407	0.7244	0.6054	0.9016
BFocal( $w_{1} = 0.2$ )	0.9848	0.8658	0.7633	0.7900	0.7384
BFocal( $w_{1} = 0.3$ )	0.9847	0.8614	0.7597	0.7927	0.7293
BFocal( $w_{1} = 0.4$ )	0.9848	0.8647	0.7618	0.7893	0.7361
BFocal( $w_{1} = 0.5$ )	0.9847	0.8687	0.7627	0.7817	0.7446

Table 3. Ablation study

View table

View in Article

Table 3. Ablation study

实验组	ResNet	ConvNeXt	DUC	PPM	aAcc/%	mAcc/%	F1/%	P/%	R/%
1	√				97.43	87.39	76.38	77.25	75.54
2		√			98.45	88.17	78.56	79.52	77.62
3		√	√		98.49	90.18	78.06	75.08	81.28
4		√		√	98.59	87.76	78.71	81.46	76.15
5		√	√	√	98.51	90.89	80.92	79.20	82.72

Table 4. The test based on the established UCrack dataset
View table
View in Article
Table 4. The test based on the established UCrack dataset
训练集大小/张 aAcc/% mAcc/% F1/% P/% R/%
959 98.49 86.95 76.54 78.59 74.60
1438 98.56 87.89 77.81 79.21 76.47
2876 98.51 90.89 80.92 79.20 82.72

Table 5. The test results on the CrackDataset dataset
View table
View in Article
Table 5. The test results on the CrackDataset dataset
训练集大小/张 aAcc/% mAcc/% F1/% P/% R/%
959 96.99 90.44 79.23 75.92 82.84
1438 98.45 91.18 78.13 73.49 83.39
2876 97.05 91.79 80.11 75.22 85.68

Table 6. The test results on training datasets with different scene coverage

View table

View in Article

Table 6. The test results on training datasets with different scene coverage

训练数据集大小/张	测试集	aAcc//%	mAcc/%	F1/%	P/%	R/%
UCrack(959)	UCrack	98.49	86.95	76.54	78.59	74.60
UCrack(959)	CrackDataset	96.99	90.44	79.23	75.92	82.84
CrackDataset(621)	UCrack	86.41	74.95	24.53	15.26	62.63
CrackDataset(621)	CrackDataset	98.21	94.77	86.10	81.82	90.86

Table 7. The test results on the dataset of models

View table

View in Article

Table 7. The test results on the dataset of models

模型	测试数据集	aAcc//%	mAcc/%	F1/%	P/%	R/%	FPS
本文模型	UCrack	98.51	90.89	80.92	79.20	82.72	2.4
	CrackDataset	97.05	91.79	80.11	75.22	85.68	2.4
FCN	UCrack	98.51	86.70	76.70	79.54	74.05	3.6
	CrackDataset	96.98	86.27	74.85	75.65	74.07	3.6
U-Net	UCrack	98.52	87.33	77.10	78.92	75.35	3.4
	CrackDataset	97.42	89.01	78.87	78.29	79.45	3.4
spent	UCrack	98.54	85.86	76.61	81.48	72.29	1.7
	CrackDataset	97.13	86.92	76.08	76.88	75.29	1.7
DeepLabv3+	UCrack	98.60	87.61	78.24	80.80	75.83	3.1
	CrackDataset	97.35	87.87	77.93	78.8	77.08	3.1

Tools

Get Citation

Copy Citation Text

Yanning LIU, Guobao ZHANG. Design and research on pavement crack segmentation based on convolutional neural network[J]. Journal of Applied Optics, 2024, 45(2): 373

Download Citation

EndNote(RIS)BibTex Plain Text

Set citation alerts for article

Save article for my favorites