Real Time Face Recognition Based Attendance System using Multi Task Cascaded Convolutional Neural Network Vrushaket Chaudhari Department of Computer Engineering, SCTR'S Pune Institute of Computer Technology, Pune, Maharashtra, India vrushu00@gmail.com Tanvesh Chavan Department of Computer Engineering, SCTR'S Pune Institute of Computer Technology, Pune, Maharashtra, India tanveshchavan2000@gmail.com Shantanu Jain Department of Computer Engineering, SCTR'S Pune Institute of Computer Technology, Pune, Maharashtra, India shantanujain18@gmail.com Prof. Priyanka Shahane Department of Computer Engineering, SCTR'S Pune Institute of Computer Technology, Pune, Maharashtra, India priyankashahane3696@gmail.com Rushikesh Chaudhari Department of Computer Engineering, SCTR'S Pune Institute of Computer Technology, Pune, Maharashtra, India rushikeshchoudhari2001@gmail.com Abstract— Facial recognition has been an important research direction in computer vision. There are countless algorithms presented in related disciplines, and the precision that may be achieved is increasing. However, the implementation of facial recognition technology is hard. In this paper, combination of facial recognition and facial recognition algorithms to build a video-based facial recognition system to efficiently and accurately mark participant attendance. Utilizing FaceNet to extract characteristics and use MTCNN to detect the image of the student for recognition. Lastly, the output is analyzed by a Support Vector Machine (SVM) that recognizes the person of interest in the image. Studies reveal that this technique still yields accurate detection results when the dependent variable has no data and the image quality is unreliable. On the self-generated data set used in this article, the accuracy of the procedure may reach 94.85%. Keywords— Face detection, Face recognition, MTCNN, FaceNet, SVM. I. INTRODUCTION Today, attendance systems are very important in companies, schools, governments, and other places where human resource management is required. The presence of fingerprints requires a queue for identification, which takes a lot of time. Injury to the finger can greatly reduce the accuracy of fingerprint recognition, and fingerprints can also be forged by others. Scanning an ID card for attendance does not verify the identity of the cardholder, which also leads to fraudulent attendance behaviour. Since iris detection is used for presence, the detection speed is slow and the time and equipment costs are high. Location-based attendance checks on mobile phones are like scanning an ID card, but cannot verify the user's identity, and the location can be spoofed. With the continuous development of machine learning and artificial intelligence technology, the methods of face recognition, facial recognition and facial feature recognition have undergone major changes. As an important biological feature, the human face has been widely used in attendance systems. Dynamic facial recognition technology eliminates the need for users to stop and wait for verification. Users simply appear as part of video surveillance and are automatically recognized by the system. Due to its real-time nature and convenience, this technology has become a hot research target for attendance systems [1]. In this work, a presence system based on multiple facial recognition is designed. To prevent users from using their photos for attendance, the system has the ability to detect blink movements. Finally, for ease of use, a user interaction interface for the attendance system is design. This project describes the different algorithms used for facial recognition system using videos that can identify a person using his facial features The following work is done in this study to address issues with face recognition systems that use video. 1. The datasets necessary for facial recognition based on video are generated. The method described in this article is used to pinpoint certain video characters. Facial recognition training models was unable to use images from open source datasets since the system needs to detect certain characters. The data set, which consists of facial images is taken from cameras without any target subjects, is used to finish training the face recognition model. 2. Take the video key frames out next. This significantly lowers the computational difficulty of video face recognition, significantly shortens the time it takes to recognize a picture, and accelerates video face recognition. Facial detection using MTCNN is applied to extract faces from the given frames. 3. Adopts a facial feature extraction method based on deep learning. Target faces in videos are recognized by his SVM classifier in this document. To avoid classifier performance degradation caused by category growth, a method is proposed to generate a binary SVM classifier by target. Additionally, the method developed in this article may successfully solve the issues of a lack of relevant sample data and surpass video quality, which are frequent in the real- world use of facial recognition algorithms. Even 2023 International Conference on Emerging Smart Computing and Informatics (ESCI) AISSMS Institute of Information Technology, Pune, India. Mar 1-3, 2023 978-1-6654-7524-2/23/$31.00 ©2023 IEEE 1