Mobile/Embedded-CV Model-2018: MobileFaceNets

Original paper: MobileFaceNets: Efficient CNNs for Accurate Real-Time Face Verification on Mobile Devices

With the development of technology, face recognition algorithms are more and more widely used in embedded terminals, but due to the limitation of computing power and storage resources of terminal equipment, the requirements for face detection and recognition models tend to be lightweightClass + high precision.Compared with the deep and wide large model, the lightweight model has the characteristics of small parameter quantity and less multiplication and addition, but at the same time, it cannot have too much loss in prediction accuracy.

In recent years, lightweight networks such as MobilenetV1, ShuffleNet and MobileNetV2 are mostly used for visual recognition tasks of mobile terminals, but due to the particularity of the face structure, these networks have not achieved satisfactory results in face recognition tasks.In response to this problem, Sheng Chen et al. of Beijing Jiaotong University proposed a lightweight network MobileFaceNet for face recognition in the paper "MobileFaceNets: Efficient CNNs for Accurate RealTime Face Verification on Mobile Devices".

As shown in the figure below, when using networks such as MobileNetV2 for face recognition, the average pooling layer gives the same weight to the Corner Unit and Center Unit of FMap-end, but in fact, for face recognition,The importance of the central element is obviously more important than that of the corner elements.Therefore, targeted optimization of the network is required.In the paper, the most important optimization is to use Global Depthwise Convolution (GDConv, global depthwise convolutional layer) instead of Global Average Pooling (GAP, global average pooling layer), because the weights of GDConv is equivalent to realizing the importance of different positionsweight factor.

References:
MobileFaceNet Model Analysis