Research on Face Detection Based on Lightweight MTCNN

In this paper, the face detection is realized by MTCNN, and the MTCNN is lightweighted. The O-Net binarization with the most parameters in the three shallow sub-networks of the MTCNN is carried out to compress the network model, the import feature map and the weight of the convolution operation are binarized so that the multiplication and accumulation operation of the convolution operation part in the original neural network is further transformed into the combination of the single-bit XNOR operation and the addition operation. In this paper, the MTCNN model before and after lightweighting are trained and tested on the WIDER FACE face detection data set respectively. The lightweight network model has a slight loss of 2.408% of the accuracy rate of training tests on the WIDER FACE face detection data set, but a 67.265% reduction in the number of parameters compared to the original model.