Overview of unconstrained low resolution face recognition III: homogeneous low resolution face recognition methods

   

        The deep learning method has been proved to be effective for general recognition tasks . However , Their complexity makes them impossible to use in real-time scenarios . therefore , There are ways to achieve real-time performance in other computer vision tasks on embedded devices , for example SqueezeNet、ShuffleNet、ShuffleNet v2、MobileNet、MobileNet v2、MobileNet v3 and VarGNet. This group of methods is to effectively solve general computer vision tasks （ Such as image recognition 、 Target detection, etc ） Proposed .

The figure above shows the multiplication and addition operations for the accuracy performance of these computer vision tasks （MADD） The benchmark , To roughly understand the location between them .MobileNetV3 It is the most efficient task in general computer vision 、 The most accurate network .

1、 Lightweight convolutional neural network for face recognition （MobileFaceNet, ShuffleFaceNet ,VarGFaceNet ）

Using technology ：a、 Use packet convolution and shuffle output channels , To reduce the number of operations , And share information between different input and output channels ;

b、 Use variable packet convolution groups to balance information retention and complexity , Point by point 1×1 Convolution , To reduce the computational complexity of depth channel and filtering ;

c、 Have low dimensional embedding before fully connecting layers , Use stepping rather than maximum pool operations to reduce complexity and retain more information directly from the data ;

d、 Use the reverse bottleneck structure to reduce the number of parameters and compress the network channel again to match the input channel .

2、 Knowledge distillation and quantitative network  

VarGFaceNet Use knowledge distillation for training , This method uses Recursive knowledge distillation and Angular distillation loss function .

a、 The selected teacher network is a ResNet Architecture , For feature vector extraction . And then in VarGFaceNet These eigenvectors are used in the loss function , To draw the feature vector generated by the lightweight network , Make it closer to the eigenvector of the teacher network .

b、 The angular distillation method uses similarity measurement , for example L2 Or cosine similarity to score the two eigenvectors .

3、 Conclusion

        The most successful unconstrained very low resolution face recognition methods are reviewed , At the same time, the limitations and advantages of each method in the latest technology are discussed . We discussed the factors that affect the accuracy and reasoning time performance , And precautions for using different training methods . We analyzed in Architecture level 、 Loss of functional design and image representation levels bridge the gap in the field Influence . For this reason , We also discuss the most important trend of deep learning convolutional neural networks as a whole , Include Capsule network 、CNN Multi branch architecture and knowledge distillation .