A Defect Detection Model for Industrial Products Based on Attention and Knowledge Distillation

Table 2

Comparative experiment of the aluminium defect dataset.

Approach	Backbone	[email protected] (%)	[email protected] (%)	Precision (%)	Recall (%)	F1	Model size (M)	FPS
SSD	VGG-16	86.58	35.2	99.8	62.5	0.77	99.76	67
Faster-rcnn	ResNet-50	92.35	42.3	73.88	97.5	0.84	523	16
YOLOv3	DarkNet-53	95.9	54.4	94.7	95.7	0.95	234	19
YOLOv5S	CSPDarkNet	97.2	56.6	96.4	96.8	0.97	26	50
Efficientdet-d3	EfficientNet-B3	93.52	50.2	91.85	90.91	0.91	14.78	20
YOLOv5X	CSPDarkNet	94.3	54.9	95.1	94.7	0.95	167	21
Centrenet	Hourglass-104	60.3	38.6	51.09	60.61	0.55	124.61	48
Retinanet	ResNet-101	96.52	46.7	94.79	93.32	0.94	144.84	20
YOLOv4	CSPDarkNet-53	94.13	59.1	83.38	96.01	0.89	245.53	29
YOLOR-P6	CSPDarkNet	97	53.1	88.3	97.8	0.93	140.88	61
YOLOX-s	Darknet-53	99.2	58.2	98.91	99.3	0.99	34.2	41
T-model	Our backbone	98.53	58.8	98.85	98.38	0.98	81.6	19
S-model	Our backbone	97.82	55.86	97.83	96.87	0.97	25.1	36