Face detection method based on lightweight cascade network

Abstract

The invention discloses a face detection method based on a lightweight cascaded neural network. The cascade network is provided with three cascade sub-networks, namely, a P-Net network, an R-Net network and an O-Net network. Each sub-network has three outputs which are face / non-face classification, face candidate box position and face feature point positioning. The method comprises the steps thatfirstly, an image pyramid is constructed so that a human face in a picture is zoomed to a proper size which can be detected by P-Net, the P-Net is used for detection, and a human face candidate windowis rapidly generated; and then, the R-Net network further filters the face candidate windows generated by the P-Net network, and refuses most of the non-face windows; and finally, the face candidatewindow generated by the R-Net network again is filtered by the O-Net network, and a final face candidate box and five face feature points are outputted. The present invention can be deployed in a resource-limited device. Compared with an MTCNN framework, the speed is increased by 25%, and resource occupation is reduced by more than 40%.

Description

technical field [0001] The invention relates to a face detection method based on a lightweight cascade network, and belongs to the technical field of face detection. Background technique [0002] In recent years, due to the high efficiency and high accuracy of deep learning algorithms, it has been increasingly used in face detection. Pang et al. designed a two-stage cascaded residual network for face detection and achieved the highest accuracy in the 2015 binocular stereo matching dataset. Qin et al. proposed a joint training method of cascaded networks to improve the accuracy of the model. Jiang et al. used the Faster-RCNN method for face detection on the WIDER face data set, and achieved good results, while speeding up the calculation speed. By designing a three-layer cascaded network, Zhang et al. made the model more accurate than 92% in the FDDB dataset. [0003] Existing face detection algorithms are usually based on large-scale neural networks, and their parameters ...

AI Technical Summary

Problems solved by technology

[0003] Existing face detection algorithms are usually based on large-scale neural networks, and their parameters (Params) usually exceed 20M, and the floating-point computing power (FLOPs) required for operation is greater than 1000M. Although these networks can achieve high accuracy, but Because they consume a lot of computing resources, when deployed in a device with limited resources, it takes more than 1 second to perform face detection and recognition, which will cause the device to freeze and even affect the normal use of the device.

Images