BeautyNet: Joint Multiscale CNN and Transfer Learning Method for Unconstrained Facial Beauty Prediction

Table 3

Comparison of network parameters using different activation functions.

Proposed network	Activation function	No. of parameters	Storage space (M)	Time (ms/frame)	Classification accuracy (%)
LightenNet	ReLU	1357664	20.9	224.57	62.79
	Sigmoid			234.64	—
	Tanh			233.90	—
	MFM			180.78	64.36
BeautyNet	ReLU	2671461	55.2	287.94	63.48
	Sigmoid			282.18	—
	Tanh			291.10	63.48
	MFM			203.68	64.84

— indicates that the model does not converge.