Skip to main content
. 2021 Sep 13;8:726943. doi: 10.3389/fcvm.2021.726943

Figure 4.

Figure 4

Different attention unit MobileNetV2 training results of 4x traditional augmentation. (A) Spatial attention (SA) accuracy curve (89.8% testset accuracy). (B) Channel attention (CA) accuracy curve (92.2% testset accuracy). (C) Accuracy curve without any attention (89.8% testset accuracy). (D), Mix attention (MA) accuracy curve (89.3% testset accuracy). The accuracy curve fluctuate more violently without any attention module and the final accuracy is the lowest.