Table 8 Comparison of the performance of different methods on the SSGD dataset.

From: Smartphone screen surface defect detection using dynamic large separable kernel attention and multi-scale feature bi-directional path aggregation network

Method

Precision (%)

Avg_Recall (%)

mAP@0.5 (%)

mAP@0.5:0.95(%)

Crack

Broken

Spot

Scratch

Light-leakage

Blot

Broken-membrane

YOLOv3

35.8

50.3

23.3

38.2

4.75

19.0

35.3

27.6

29.5

12.8

YOLOv5

44.1

59.0

44.7

60.7

17.3

10.0

34.5

35.2

38.6

17.6

YOLOv6

41.9

59.0

42.5

55.1

10.7

1.26

30.5

35.9

34.4

15.3

RTDETR

23.0

57.8

49.6

50.6

21.7

29.2

33.9

40.9

38.0

17.8

YOLOv8

45.8

63.5

43.8

59.0

25.0

10.0

48.9

41.6

41.2

17.6

YOLOv9

41.5

63.8

43.8

55.8

13.0

9.72

23.8

41.6

35.9

17.3

YOLOv10

39.7

59.9

45.8

51.7

17.9

11.0

35.2

34.4

37.3

16.7

YOLOv11

40.3

44.0

48.6

54.5

82.3

31.3

45.9

42.6

40.1

17.8

YOLOv12

53.0

57.7

53.3

50.4

31.3

0

33.1

33.0

34.9

15.4

DY-YOLO

49.0

68.1

46.8

58.8

11.8

46.2

41.1

46.2

46.0

20.2