A Multi-Task Framework for Facial Attributes Classification through End-to-End Face Parsing and Deep Convolutional Neural Networks

. 2020 Jan 7;20(2):328. doi: 10.3390/s20020328

Algorithm 1 proposed RAG-MCFP-DCNNs algorithm

Input:M $_{t r a i n}$ = {(I $_{i}$ ,T $_{i}$ )}

_{i = 1}^{j}

, M $_{t e s t}$
where the DCNNs model is trained through training data represented as M $_{t r a i n}$ and tested through M $_{t e s t}$ . The input training image is represented as I and the ground truth data is T(i,j)∈ {1,2,3,4,5,6,7}.
a: Face parsing part:
Step a.1: Training a face parsing model DCNNs through training images and class labels.
Step a.2: Using the probabilistic classification strategy and producing PMAPs for each semantic class, represented as:
PMAP

_{s k i n}

, PMAP

_{m o u t h}

, PMAP

_{e y e s}

, PMAP

_{n o s e}

, PMAP

_{h a i r}

, PMAP

_{b a c k}

, and PMAP

_{e y e b r o w}

b. race, age and gender classification part:
Training a second DCNNs for each demographic class (race, age, and gender) by extracting infomration from PMAPs of the corresponding classes such that;
if race classification:

f =

PMAP

_{s k i n}

+ PMAP

_{m o u t h}

+ PMAP

_{e y e s}

+ PMAP

_{n o s e}

+ PMAP

_{h a i r}

+ PMAP

_{b r o w s}

Else if age classification:

f =

PMAP

_{s k i n}

+ PMAP

_{m o u t h}

+ PMAP

_{e y e s}

+ PMAP

_{n o s e}

+ PMAP

_{b r o w s}

Else if gender recognition:

f =

PMAP

_{s k i n}

+ PMAP

_{e y e s}

+ PMAP

_{b r o w s}

+ PMAP

_{n o s e}

+ PMAP

_{m o u t h}

where f is the feature vector.
Output: estimated race, age and gender.