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

Table 7 Comparison of the MSCNet with other advanced backbone models on the PCB dataset

Method	mAP(%)	Speed(FPS)	Size(MB)	FLOPs(G)	Param.(M)	\(AP_{50:95}\)	\(AP_{50}\)	\(AP_{75}\)	\(AP_{s}\)	\(AP_{m}\)	\(AP_{l}\)
ResNet50+SFPN	75.59	81.88	119.72	50.98	29.78	0.342	0.750	0.209	0.202	0.342	0.275
EfficientNet+SFPN	76.86	78.56	33.62	11.60	8.23	0.326	0.757	0.208	0.050	0.328	0.280
GhostNet+SFPN	77.57	82.38	20.88	7.98	5.00	0.333	0.771	0.211	0.151	0.341	0.267
MobileNetV2+SFPN	81.33	86.44	18.71	9.68	4.54	0.366	0.805	0.246	0.505	0.368	0.310
InceptionV3+SFPN	84.93	67.48	111.66	49.30	27.69	0.386	0.836	0.265	0.050	0.392	0.323
Xception+SFPN	86.05	82.17	107.33	57.39	26.71	0.373	0.854	0.230	0.151	0.373	0.304
Densent121+SFPN	86.11	65.35	45.94	38.25	11.20	0.385	0.856	0.263	0.050	0.395	0.337
DarkNet+SFPN	93.64	84.66	179.72	82.89	44.78	0.431	0.930	0.232	0.176	0.433	0.400
CSPDarknet53+SFPN	94.58	97.18	124.00	60.85	30.82	0.435	0.931	0.316	0.184	0.426	0.417
VGG16+SFPN	94.90	100.18	72.58	168.42	18.09	0.439	0.945	0.327	0.252	0.432	0.431
MSCNet+SFPN(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