Deeply Learned Classifiers for Age and Gender Predictions of Unfiltered Faces

Table 1

The details of CNN architecture for the proposed method.

Layer type	Output size	Filter size/stride
Input image	227 × 227 × 3	—
CONV1	56 × 56 × 96	7 × 7/4 × 4
ACT	56 × 56 × 96	—
BN	56 × 56 × 96	—
Maxpool	28 × 28 × 96	3 × 3/2 × 2
Dropout	28 × 28 × 96	—
CONV2	28 × 28 × 256	5 × 5
ACT	28 × 28 × 256	—
BN	28 × 28 × 256	—
Maxpool	14 × 14 × 256	3 × 3
Dropout	14 × 14 × 256	—
CONV3	14 × 14 × 384	3 × 3
ACT	14 × 14 × 384	—
BN	14 × 14 × 384	—
Maxpool	7 × 7 × 384	3 × 3
Dropout	7 × 7 × 384	—
CONV4	7 × 7 × 384	3 × 3
ACT	7 × 7 × 256	—
BN	7 × 7 × 256	—
Maxpool	1 × 1 × 256	3 × 3
Dropout	1 × 1 × 256	—
FC1	512	—
ACT	512	—
BN	512	—
Dropout	512	—
FC2	2 or 8	—