YOLO-MSD: a robust industrial surface defect detection model via multi-scale feature fusion

Table 8 Comparison with advanced models based on the open dataset PCB

Method	mAP(%)	Speed(FPS)	Size(MB)	FLOPs(G)	Param.(M)	\(AP_{50:95}\)	\(AP_{50}\)	\(AP_{75}\)	\(AP_{s}\)	\(AP_{m}\)	\(AP_{l}\)
Faster-RCNN (VGG16)	44.38	42.17	547.16	226.41	119.65	0.121	0.441	0.037	0.006	0.105	0291
YOLOv3	84.29	70.70	246.99	99.30	61.60	0.312	0.803	0.113	0.151	0.314	0.336
YOLOv4	86.38	68.69	256.97	90.58	64.03	0.344	0.855	0.155	0.084	0.341	0.397
YOLOv4-Tiny	75.40	137.16	23.74	10.35	5.89	0.271	0.744	0.094	0.101	0.275	0.256
YOLOv5-X	10.64	69.69	346.35	130.75	86.34	0.039	0.109	0.010	0.000	0.039	0.049
YOLOX-X	91.48	59.56	397.62	180.18	99.10	0.432	0.907	0.305	0.252	0.426	0.455
YOLOX-L	85.06	71.24	217.89	99.41	54.22	0.382	0.837	0.240	0.151	0.383	0.365
YOLOX-M	90.54	68.21	102.15	47.05	25.33	0.419	0.893	0.310	0.202	0.420	0.405
YOLOX-S	87.60	75.24	36.47	17.05	8.96	0.401	0.864	0.281	0.252	0.399	0.418
YOLOX-Tiny	82.50	81.15	20.83	9.69	5.05	0.374	0.820	0.237	0.151	0.379	0.332
YOLOv7-X	90.13	64.77	284.48	120.66	70.91	0.399	0.888	0.275	0.252	0.398	0.451
YOLOv7-L	40.00	68.00	149.82	67.10	37.27	0.143	0.400	0.060	0.000	0.155	0.070
YOLO-MSD-X	95.86	62.61	317.33	135.59	79.18	0.443	0.950	0.299	0.151	0.441	0.391
YOLO-MSD-L	96.67	70.13	192.93	82.85	48.08	0.429	0.954	0.292	0.202	0.425	0.460
YOLO-MSD-M	94.46	74.22	99.43	43.04	17.58	0.443	0.936	0.319	0.202	0.443	0.435
YOLO-MSD-S	94.15	75.72	36.79	16.16	9.05	0.424	0.936	0.282	0.252	0.419	0.401
YOLO-MSD-Tiny	90.64	78.78	18.24	8.32	4.41	0.408	0.900	0.263	0.151	0.411	0.323