Table 4 The influence of hyperparameters in MFE on model detection accuracy.

From: Object detection model design for tiny road surface damage

Model No

Residual connection

MFE_PCR

P(%)

R(%)

mAP50(%)

mAP50:95(%)

F1

c1:c2:c3

cin:cout

1

 

2:1:1

2:1

71.2

61.6

65.8

34.9

66.1

2

2:1:1

2:1

71.5

62.1

66.1

35.0

66.5

3

1:1:1

2:1

71.6

61.2

65.6

34.6

66.0

4

3:1:1

2:1

71.4

62.0

66.1

35.3

66.4

5

4:1:1

2:1

71.4

61.8

66.2

35.1

66.3

6

5:1:1

2:1

71.4

62.9

66.5

35.4

66.9

7

6:1:1

2:1

70.3

63.3

66.1

35.1

66.6

8

5:1:1

3:1

69.8

61.1

65.3

34.6

65.2

9

5:1:1

3:2

72.1

61.9

66.1

35.3

66.6

  1. In the table, “” indicates the use of residual connections in MFE_SCR and MFE_PCR, “c1:c2:c3” represents the ratio of the number of output channels of the parallel convolutional layer in MFE_PCR, and “cin:cout” represents the input–output ratio of the first 1 × 1 convolution in MFE_PCR.