Lightweight distraction judging method based on deep learning face recognition

Abstract

The invention discloses a lightweight distraction judging method based on deep learning face recognition, which is used for processing video stream data and comprises the following steps of: carrying out face detection on a key frame, then carrying out distraction recognition on a detected face region, and finally obtaining a distraction state recognition result of the frame; the method mainly comprises two modules: a face detection algorithm based on ResNet10-SSD, and a distracting recognition algorithm based on MobileNet+GRU. The face detection module continues to use the SSD, the distracting recognition module adopts depth separable convolution of MobileNet to build a feature extractor, a CBAM structure retaining identical mapping is added, and face key point positioning and head posture estimation are adopted as auxiliary data for additional supervision; due to the fact that ResNet10 and other lightweight backbone networks are adopted and MobileNet is used for acceleration, the lightweight facial distraction recognition method gives consideration to precision and speed, can be deployed on various mobile devices, has the distraction recognition accuracy reaching 90%, and has high practical value in practical application scenes.

Description

technical field [0001] The invention relates to the field of deep learning image recognition and analysis, in particular to a light-weight mind-wandering discrimination method based on deep learning facial recognition. Background technique [0002] The development of information technology has brought great convenience to people's lives, especially for the online education industry. The form of live video allows students to complete classroom learning at home, which is even more important in a society under the epidemic. In order to improve the quality of students' lectures and allow teachers to get timely feedback, it is necessary to intelligently identify whether students are distracted during the live broadcast process, and summarize the results for teachers' reference and processing. For the video stream captured by the camera of the student's listening device, it is necessary to detect the location of the student's face in the key frame, and then perform distraction rec...

AI Technical Summary

Problems solved by technology

However, existing datasets and models are limited to facial expression recognition tasks that reflect basic emotions such as anger, high sex, sadness, and surprise. The DAiSEE mind-wandering recognition dataset proposed by Indian scholars is also not applicable due to racial differences and video quality problems. In domestic business scenarios, there is no model that shows good performance on the DAiSEE dataset

Therefore, in the field of facial distraction recognition, both data sets and network models are very scarce, which brings certain difficulties and challenges

Images