Face Detection Based on Improved Multitask Convolution Network

Abstract

Aiming at the problem of inaccurate location caused by missing information after feature extraction in traditional multitask cascade convolution network, feature pyramid structure is used to strengthen the context connection of images; aiming at the inaccuracy of face detection caused by difficult samples, an new combined loss function is applied; and a improved multitask convolution network is proposed based on softNMS (Non maximum suppression). The network shows excellent performance on public data sets self-made data sets, and has higher accuracy than traditional MTCNN (Multi-task convolutional neural network).