1.Research Background
The most widely used MTL convolutionalneural network is empirically designed by sharing all of the convolutionallayers and splitting at the fully connected layers for task-specific losses.
2. (Overview)MTCN+NTCCA
Propose a novel multi-attribute tensorcorrelation neural network (MTCN) for face attribute prediction.
Project the features of the C9 layers ofthe fine-tuned subnetworks into a highly correlated space by using a noveltensor correlation analysis algorithm (NTCCA). The final face attributeprediction is made based on the correlation matrix.
3. Method
3.1 Low-level Feature Sharing for Face Attributes
sharing information in low-level featurelayers
3.2 Differentiationand Correlation in High-level Layers
From the third convolutional layer, wesplit the network into multi-subnetworks.
each of our subnetworks seeks usefulinformation from the other networks in the same layers to enhance itself.
3.3 Multi-attributeTensor Correlation Learning Framework
Not fully consider the specific degrees ofcorrelation among the face attributes.
NTCCA explore the detailed correlations among the high-level features of thesesubnetworks.
TCCA, aims to directly maximize thecorrelation between the canonical variables of all views;
NTCCA maximizes the correlation of all ofthe feature maps in C9 for an image.
Ci expresses the cross-entropy loss of thei-th attribute,
Φ(·) denotes a regularization term topenalize the complexity of the weights, γ > 0 is a regularization parameter.
3.4 Novel TensorCanonical Correlation Analysis
4. EXPERIMENTS
Attributes can be divided into three majorcategories based on the results
(1)relatively easy to predict(exceed 95%)
correlated with one or more otherattributes,
MTCN excavates correlations in differentlevels.
(2) less than 80%
easily influenced by shooting angle andpose
few attributes are highly related
(3)the attributes in III can enhance thefeatures of the attributes in I,
while those of III benefit less from thoseof I.
5. CONCLUSIONS
Propose MTCN and NTCCA
Fully explores the correlations atdifferent levels,
including sharing information in thelow-level feature layers,
splitting that in the high-level featurelayers
while extracting related information fromother subnetworks to enhance its features
excavating the correlation of high-levelfeatures
6. References
[1] Mingxing Duan, Kenli Li, Qi Tian. A NovelMulti-Task Tensor Correlation Neural Network for Facial Attribute Prediction.In ACM MM, 2018.