Faster R-CNN based face detection for streaming media

Abstract:
Face recognition technology has been studied since the 1960s, and face detection is the most critical step in face recognition. In the early days of computer vision, face detection was a classical problem that deeply studied machine vision. It has important applications in security surveillance, human-witness comparison, human-computer interaction, social networks, etc. However, it is slightly lacking in the accuracy of detection. The current streaming media is growing, and the characters of movies and TV shows are gradually being replaced. Many people want to know the chain in the videos but do not know the information related to the people playing them. To realize the detection work of face recognition in streaming media, this paper adopts the Faster-RCNN model and adopts the method of fusing multiple loss functions in the calculation of object loss, which reduces the loss of feature object framing caused by the overlap rate and the loss of not including feature objects, to solve the problem of missing the overlapping faces and thus improve the accuracy of face detection. The model in this paper is compared with the previously used Faster-RCNN model to determine the stability and reliability of the model. Finally, the trained face detection model is put into the face detection of streaming media to frame the face accurately and provide a convenient way for the subsequent face recognition development. The overall recognition accuracy of this experiment reaches 93.2%, and the reliability of the accuracy is high throughout the results.