Compression algorithm for live face recognition model based on depth-separable convolution

In this paper, we present a new compression algorithm for live face recognition models. Firstly, a VGG11 convolutional neural network is improved based on depth-separable convolution. The depth-separable convolution layer consists of a depth convolution layer and a point-by-point convolution layer. The depth convolution is used for filtering, acting on each channel of the input. And the point-by-point convolution is used for transforming the channels, acting on the output feature mapping of the deep convolution. Compared with the original model, although the accuracy of the improved model decreases by about 1%, the model size reduces to about 1/5 of the original size. Secondly, since the float16 quantization approximates continuously valued floating-point model weights to a finite number of discrete values with a low loss of accuracy, the model volume is compressed by half with little impact on model accuracy through float16 quantization. The proposed method can compress the original face live recognition model by 90.94% and reduce the recognition time by 83.46% with 1.07% accuracy loss.