Table 3 The influence of different feature extraction strategies at each stage on model detection accuracy.

From: Object detection model design for tiny road surface damage

Multi-strategy feature extraction

P(%)

R(%)

mAP50(%)

mAP50:95(%)

F1

Stage 1

Stage 2

Stage 3

Stage 4

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

71.9

59.7

65.2

34.5

65.2

MFE_PCR

(1,5,9)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

70.8

60.6

65.2

34.4

65.3

MFE_PCR

(1,3,5)

MFE_PCR

(1,2,3)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

69.5

60.9

65.0

34.4

64.9

MFE_PCR

(1,2,3)

MFE_PCR

(1,2,3)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

70.9

61.5

65.6

34.5

65.9

MFE_PCR

(1,2,3)

MFE_PCR

(1,2,3)

MFE_PCR

(1,2,3)

MFE_PCR

(1,2,3)

70.7

60.0

64.9

34.0

64.9

MFE_PCR

(1,2,3)

MFE_PCR

(1,2,3)

MFE_PCR

(1,2,3)

MFE_PCR

(1,3,5)

69.8

61.3

64.7

34.2

65.3

MFE_SCR

MFE_PCR

(1,2,3)

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

71.5

62.1

66.1

35.0

66.5

MFE_SCR

MFE_SCR

MFE_PCR

(1,3,5)

MFE_PCR

(1,3,5)

69.7

61.0

65.1

34.6

65.1

MFE_SCR

MFE_SCR

MFE_SCR

MFE_SCR

68.8

59.2

63.5

33.7

63.6

  1. In the table, “MFE_PCR (1,5,9)” represents the dilation rate of the parallel convolutional layers in MFE_PCR is 1,5,9 respectively, and “MFE_PCR (1,2,3)” represents the dilation rate is 1,2,3 respectively.