Face Detection Based on Cascaded Multivariate Branches Convolutional Neural Networks

Abstract To improve MTCNN’s face detection accuracy, this paper proposes multivariate branches block (MBB) and dilated convolutional block (DCB), and we design a cascaded multivariate branch convolutional neural network model(MB-MTCNN) based on these two structural blocks. First, the model consists of three subnetworks, and we replace the first convolutional block of each subnetwork with a dilated convolutional block. Second, replace the second convolutional block of the second and third subnetworks with MBB. In addition, we add a BN layer after each convolutional layer to accelerate the training speed of the model. Finally, we introduce a structural re-parameterization method to reduce the model’s overhead in inference. Experimental results show that our model increases the True Positive Rate (TPR) by 4.48% on the FDDB datasets, improves the average accuracy by 8.08%, 8.65%, and 10.29% on the three levels of validation sets of Wider Face datasets, 8.31% and 2.08% on the AFW and CelebA datasets, respectively. Moreover, based on structural re-parameterization, the model parameters during inference are 14.68% lower than training. This model is suitable for scenarios that require accurate detection of faces, such as crowded stations and traffic arteries.