Table 5 Per-class performance on NEU-DET (vs. best baseline: +3.7% \(\hbox {mAP}_{50}\); Cr +43.4%, In +4.46%, Ps +3.99%).

From: Multiscale diffusion-enhanced attention network for steel surface defect detection in Polysilicon Production

Datasets

Method

Cr

In

Pa

Ps

Rs

Sc

\(\hbox {mAP}_{50}\)

NEU-DET

YOLOv5s

39.8

72.2

96.6

87.7

71.0

94

76.9

YOLOv8s

30.3

82.9

94.3

85.4

61.0

90.8

75.2

YOLOv9s

39

76.2

95.6

90.2

68.8

95.8

77.6

YOLOv10s

32

70.9

94.4

77.6

68

81.4

70.7

YOLOv11s

39.5

72.1

95

88

66.4

90.8

75

FFDDNet15

41

75.2

94.3

87.1

71.5

92.0

76.8

Ours

58.8

86.6

93.6

93.8

68.5

92.7

80.5

Back to article page