Face and face shielding object detecting method based on multitask deep learning

Abstract

The invention discloses a face and face shielding object detecting method based on multitask deep learning, wherein the method comprises the steps of 1, establishing a multitask face detail detecting network, wherein the multitask face detail detecting network is composed of three cascaded subnetworks, namely an F-Net, an O-Net and a C-Net, wherein the F-Net is used for detecting an appropriate position of a face and supplies candidate areas of the face position for later two network grades; the O-Net is used for further determining a credibility degree of the candidate area based on the detecting result of the F-Net and performs correction on face bbox and is used for detecting whether a shielding object exists in the face image and determining the position of the shielding object; and the C-Net is used for further correcting the detecting results of the F-Net and the O-Net; and 2, detecting the face details in a picture based on the established multitask face detail detecting network, and obtaining a shielding object detecting result for representing whether the shielding object exists on the face, the type of the shielding object and a position detecting result. The face and face shielding object detecting method can detect position of the shielding object and furthermore realizes technical effects of replacing the shielding object and applying in scenes such as AR.

Description

technical field [0001] The present invention relates to the field of human face occlusion detection, in particular to a method for detecting human faces and human face occluders based on multi-task deep learning. Background technique [0002] Multi-task learning is an inductive transfer mechanism with the fundamental goal of improving generalization performance. Multi-task learning improves generalization through domain-specific information in related task training signals, and utilizes shared representations to learn multiple tasks in parallel training. [0003] figure 1 Shows 4 independent neural networks, each network is a function with only one output for the same input. Error backpropagation is applied to these networks to train each network individually. Since these networks are not connected to each other, the features learned by one network cannot help the learning of the other network, which can be called single-task learning. (STL, Single Task Learning). [000...

AI Technical Summary

Problems solved by technology

[0017] The present invention provides a face and face occlusion detection method based on multi-task deep learning, which solves the problem that the current face occlusion detection can only judge whether it is occluded, but cannot detect the type and occlusion area of ​​common occlusions, and the algorithm The problem of poor real-time performance, and the ability to detect the position of occluders, can be further applied to the technical effects of occluder replacement, AR and other scenarios

Images