Face detection based on receptive field enhanced multi-task cascaded convolutional neural networks

Abstract

With the continuous development of deep learning, face detection methods have made the greatest progress. For real-time detection, cascade CNN based on the lightweight model is still the dominant structure that predicts face in a coarse-to-fine manner with strong generalization ability. Compared to other methods, it is not required for a fixed size of the input. However, MTCNN still has poor performance in detecting tiny targets. To improve model generalization ability, we propose a Receptive Field Enhanced Multi-Task Cascaded CNN. This network takes advantage of the Inception-V2 block and receptive field block to enhance the feature discriminability and robustness for small targets. The experimental results show that the performance of our network is improved by 1.08% on the AFW, 2.84% on the PASCAL FACE, 1.31% on the FDDB, and 2.3%, 2.1%, and 6.6% on the three sub-datasets of the WIDER FACE benchmark in comparison with MTCNN respectively. Furthermore, our structure uses 16% fewer parameters.