Table 7 Ablation experiments on components of the feature fusion module.

From: A novel approach to palm vein image segmentation combining multi-scale convolution and swin-transformer networks

Experimental group

OA

Dice

IoU (%)

Number of parameters (M)

FLOPs(G)

Complete feature fusion module

97.8%

94.5%

89.4%

7.63

12.5

Removal of multi-scale convolution

95.2%

91.1%

85.3%

6.98

10.2

Removal of self-attention mechanism

96.0%

92.3%

86.7%

7.10

11.8

Removal of refinement layer

96.5%

93.0%

87.2%

6.85

9.7

Only channel concatenation retained

93.8%

89.5%

83.1%

6.40

8.5

Back to article page