Face recognition in an intelligent door lock with ResNet model based on deep learning

Internet of Things is considerably increasing the levels of convenience at homes. The smart door lock is an entry product for smart homes. This work used Raspberry Pi, because of its low cost, as the main control board to apply face recognition technology to a door lock. The installation of the control sensing module with the GPIO expansion function of Raspberry Pi also improved the antitheft mechanism of the door lock. For ease of use, a mobile application (hereafter, app) was developed for users to upload their face images for processing. The app sends the images to Firebase and then the program downloads the images and captures the face as a training set. The face detection system was designed on the basis of machine learning and equipped with a Haar built-in OpenCV graphics recognition program. The system used four training methods: convolutional neural network, VGG-16, VGG-19, and ResNet50. After the training process, the program could recognize the user’s face to open the door lock. A prototype was constructed that could control the door lock and the antitheft system and stream real-time images from the camera to the app.

A Modified Vector Recovery Index

In this paper, we show that averaging of the Vector Recovery Index (VRI) score for a test involving many images is not accurate and leads to bias. We demonstrate that the higher the difference in primitive count between the data files in an experiment, the higher the bias in calculating the VRI. Normalizing VRI scores is proposed to remove the bias and to get VRI scores that precisely reflects the performance based on images under scrutiny. Empirical performance evaluation on three datasets from the arc segmentation contests attached to International Workshops on Graphics Recognition 2005, 2009, and 2011 shows that the proposed normalization score provides accurate and realistic performance results than the unweighted average of VRI scores. The results based on the modified VRI score show that the vectorisation methods have lower performance than was usually thought.