Video Indexing Through QR Code of Human Faces Using MTCNN Algorithm

This paper proposes an idea to generate QR bar codes of facial images from the keyframes of a video. Using a person's face as a cue for video indexing has several different approaches. Still, all the methods are not efficient in detecting a face from video because of directional changes of the face, the brightness of the face image, illuminations of face images, etc. Also, if there is a method for efficient detection, there is a lack of compact representation of faces. A compact representation is a linear bar code with much information loss during horizontal scanning of face images. Also, face detection accuracy is an issue for the system's overall accuracy. To overcome it, this paper uses a QR bar code to represent face images. This method uses Multi-Task Cascaded Convolutional Neural Networks (MTCNN) for keyframe extraction from videos with faces available at keyframes. After keyframe detection, a QR bar code is generated from this keyframe for video indexing and recognition purposes. This technique is helpful for security purposes, Human activity recognition purposes, video surveillance, video description of communication channels, etc. This method has shown better results for indexing videos using a QR bar code with MTCNN than the Viola-Jones method for face detection.