Enhancing University Security: A Machine Learning and IoT Driven Face Recognition System for Surveillance and Attendance

Abstract Convolution Neural Networks (CNN) doesn't need any techniques to implement feature selection, extraction and learns the data based on the optimizer, layers, increasing or decreasing learning rate. For design of face recognition model we can directly use detection model Convolution Neural Networks (provided by GPU) pre-trained or HOG for faster results without any training. The images in the dataset are cropped based on the region of interest where the image gets saved only if it detects eye and the embeddings. The dataset is then encoded into NumPy arrays and stored in a pickle file for prediction. The encodings are used to detect faces over image, a live video and over a video file. To train a network from scratch we need to train a lot of images so using a pre-trained model with fine tuned weights reduces training time and images are convoluted and trained as per the architecture. To run this prototype accurately we can use CNN as a detection model to recognize images as it are accurate in prediction. Design of the Face recognition system can also be done by model training. This technique helps us to increase data if required depending on performance or hyper parameter tuning to increase the efficiency of model accuracy. We used Support Vector Machine (SVM) algorithm where we have sent cropped raw images long with wavelet transformation to train the classifier. So that the model will learn the global features in the image and can accurately predict the outcome with the accuracy of 99.63%.