Table 2 Ablation experiments to verify the effect of different methods and modules on model recognition accuracy.

Approach	MMEW		CASME II		SMIC		SAMM
Approach	UAR	UF1	UAR	UF1	UAR	UF1	UAR	UF1
ShuffleNetV2 without pre-trained	0.6252	0.6384	0.6336	0.6497	0.6180	0.6104	0.6149	0.6214
ShuffleNetV2 without SVM classifier	0.6222	0.6392	0.6207	0.6291	0.6038	0.5918	0.5925	0.5899
ShuffleNetV2	0.6524	0.6621	0.6647	0.6583	0.6394	0.6451	0.6712	0.6592
ShuffleNetV2 with CBAM	0.6684	0.6719	0.6739	0.6619	0.6581	0.6610	0.6802	0.6694
ShuffleNetV2 with improved ViT	0.6894	0.7251	0.6925	0.7008	0.7159	0.7141	0.6684	0.7410
MobileNetV2 without pre-trained	0.5882	0.5741	0.5741	0.5726	0.6150	0.6021	0.5632	0.6131
MobileNetV2 without SVM classifier	0.6091	0.6032	0.5744	0.5951	0.5886	0.6125	0.5869	0.6262
MobileNetV2⁴²	0.6448	0.6684	0.6358	0.6229	0.6768	0.6692	0.6328	0.6517
ResNet50 without pre-trained	0.6427	0.6231	0.6364	0.6253	0.6331	0.6480	0.5905	0.6059
ResNet50 without SVM classifier	0.6438	0.6571	0.6296	0.6308	0.6018	0.6319	0.6246	0.6581
ResNet50	0.6793	0.6782	0.6743	0.6682	0.6841	0.6621	0.6599	0.6827
DeIT without pre-trained	0.6614	0.6582	0.6479	0.6289	0.6617	0.6696	0.6782	0.6573
DeIT without SVM classifier*	0.6561	0.6428	0.6675	0.6708	0.6825	0.6843	0.6638	0.6592
DeiT*³⁹	0.6982	0.6820	0.6815	0.7008	0.6963	0.6921	0.7108	0.7034

Quick links

Search