Source-linked AI summary
ClassVision: AI-Powered Classroom Attendance System
Ankit Kumar Aggarwal, Veerabhadra Rao Marellapudi, Ovadia Sutton, Youshan Zhang
TL;DR
Manual attendance is time-consuming and difficult to manage in educational settings. ClassVision combines face detection, face recognition, real-time image processing, and a user-friendly interface to automate classroom attendance. RetinaFace with face-recognition verification and 50x50-pixel cropped embeddings produced the most promising reported results.
Problem
Traditional attendance processes are human-intensive and time-consuming, motivating automated attendance tracking in educational settings.
Method
ClassVision uses varied-angle training images, 50x50-pixel ROI preprocessing, RetinaFace detection, face-recognition embeddings, and a web-based HCI interface.
Results
RetinaFace detected nearly all faces in the test images, while its combination with face recognition produced remarkable classroom group-image results.
Takeaways & Limitations
The framework supports classroom attendance from group images without requiring students to stand closely in front of the camera.
Abstract
from arXiv · showhide
Students and working professionals have to go through the attendance process every day. Traditional methods of marking attendance using pen and paper or online platforms are human-intensive and time-consuming. To address the challenges in manual attendance processes, this research explores the use of face detection (FD) and face recognition (FR) technology to automate the attendance process, particularly in educational settings, and build a ClassVision course attendance system. We also propose an automated attendance system featuring a human-computer interaction (HCI) and user-friendly web interface that utilizes real-time image processing to identify and recognize students in classrooms and automatically record their attendance. We identified RetinaFace as the best face detection model, and when combined with Face Recognition for verification, it provided the most promising results with a cropped embedding of 50x50 pixels.
I. INTRODUCTION
ClassVision addresses time-consuming, error-prone, and disruptive attendance practices by combining facial detection, recognition, and a user-friendly interface for classroom attendance. The framework uses varied training views, ROI preprocessing, and classroom image analysis to automate attendance recording.
- Traditional attendance methods consume time, introduce errors, and can enable impersonation, while paper records create security and retrieval problems.
- Facial-recognition attendance requires attention to privacy and algorithm-bias concerns; the study therefore collected classroom test images from volunteers in specific rooms.
- ClassVision follows five stages: collect 112x112 training images, extract ROI and embeddings, detect test-image faces, generate embeddings, and compare embeddings.
- The proposed system includes a human-computer interaction interface for automatic course attendance in classroom settings.
- ClassVision captures student faces from front, left, right, up, and down angles to support retrieving names from classroom images.
- The framework extracts a 50x50-pixel facial ROI, then combines RetinaFace detection with Python face recognition for attendance verification.
II. RELATED WORK
Related work traces attendance automation from biometric security and face-detection research toward more efficient identification systems. Prior studies address recognition accuracy, lightweight architectures, and fraud reduction.
- Fingerprint and iris biometric systems improve security but are expensive and require more time to capture attendance.
- Face detection provides a method for identifying authorized individuals while reducing the risk of fraudulent attendance.
B. Face Recognition
The paper situates ClassVision among face-recognition and attendance systems, then describes its consent-based image collection and classroom testing setup. The dataset combines student training images with group photographs captured under varied conditions.
- Prior attendance systems used facial recognition, PCA, CNNs, IoT, and edge computing, while some required close camera proximity or slower processing.
- ClassVision uses 50x50-pixel facial regions to create embeddings intended to label students across classroom distances and sizes.
- The study collected a ground-truth dataset from multiple students and divided collection into student consent, training, and testing phases.
- The training process captured 100 images for each of 36 students, producing 3,600 images in total.
- Testing continued for up to eight weeks and produced 69 classroom group photographs covering varied poses, backgrounds, lighting, and facial expressions.
- The initial research phase used laptop cameras, while specific high-resolution cameras were reserved for future exploration.
IV. PROPOSED METHOD
ClassVision’s proposed method combines five ordered stages to generate embeddings from training images and recognize students in classroom group images. The design aims to capture attendance without manual intervention or lecture disruption.
- The method comprises one-time training-image collection, 50x50 ROI extraction and resizing, test-image face detection, embedding generation, and image recognition.
- The process emphasizes collecting suitable training data and applying specified preprocessing operations to produce quality embeddings.
- ClassVision is designed to capture attendance for groups of students without manual intervention or disruption to ongoing lectures.
A. Face Detection
ClassVision combines RetinaFace for facial localization with face-recognition processing to identify classroom faces. The workflow depends on clean training images to reduce confusion from incomplete or biased facial regions.
- A. Face Detection: RetinaFace was selected as the face-detection model for localizing faces across classroom images.It performs pixel-wise face localization across multiple face scales.
- A. Face Detection: Face Recognition Python identifies facial features after facial images are extracted, cropped, resized, and often converted to grayscale.
- A. Face Detection: Verification compares a captured face with one reference image, whereas identification compares it against multiple user images to determine the matching identity.
- A. Face Detection: Clean training images are necessary because multiple faces, partial faces, empty backgrounds, caps, and thick beards can produce biased or confusing regions of interest.The authors manually remove problematic images before extracting the ROI.
D. Region of Interest (ROI)
The ROI stage extracts and standardizes facial regions from training images before embedding generation. ClassVision uses a 50x50 ROI and stores generated embeddings with student names for later predictions.
- D. Region of Interest (ROI): RetinaFace identifies facial landmark regions, which are extracted and resized from the 112x112 training images during ROI processing.
- D. Region of Interest (ROI): Processed images are traversed from student-specific subfolders, renamed, and copied into a single destination directory during ROI preparation.
- D. Region of Interest (ROI): 50x50 pixels is the proposed optimal ROI dimension for the cropped training face images.
- D. Region of Interest (ROI): The load or generate embeddings function extracts embeddings for training images and manages the corresponding student names.It supports reusing previously generated embeddings to improve prediction execution time.
- D. Region of Interest (ROI): The function represents each training and test image with facial embeddings, using i for a training sample and j for a test sample.
F. Face Detection and Face Recognition
ClassVision detects faces in classroom images, generates embeddings, and compares them with known embeddings to identify students. Its interface lets instructors verify and manage the resulting attendance records.
- F. Face Detection and Face Recognition: RetinaFace detects facial landmarks in classroom images before detected faces are converted for face-recognition processing.
- F. Face Detection and Face Recognition: Face-recognition embeddings from detected faces are compared with known face embeddings using Euclidean distance to find the closest match.
- F. Face Detection and Face Recognition: Successful matches receive a rectangular bounding box and an identifying label over the corresponding face.
- F. Face Detection and Face Recognition: The web portal allows instructors to add courses, view attendance by day, select course images, and review or override attendance statuses.
- F. Face Detection and Face Recognition: Adding a course requires loading a CSV file containing the names of the class students.
2) Capture Attendance:
The capture-attendance interface lets users select classroom images, run face detection, and verify named faces against an attendance table. Records can then be modified, saved, and downloaded as CSV files.
- 2) Capture Attendance:: Users can select an image, perform face detection, and take attendance from the capture attendance screen.
- 2) Capture Attendance:: Attendance records can be modified, updated, and saved before downloading the revised data.
- 2) Capture Attendance:: Downloaded attendance files follow the naming convention CourseCode-mm-ddyyyy.csv.The example given is AIM 5004-04-25-2024.csv.
- 2) Capture Attendance:: The evaluation used classroom images containing 2 to 27 students and manually calculated detection and recognition accuracy.
A. Detection
This section evaluates RetinaFace detection and face recognition for classroom group images, defining accuracy using detected or correctly identified faces relative to available faces.
- Detection: RetinaFace can identify nearly all faces present in classroom test images.The paper illustrates this result with a classroom group image in Fig. 8.
- Recognition: Face recognition coupled with RetinaFace provides remarkable results for classroom group images.The paper illustrates the recognition output with a group image in Fig. 9.
- Recognition: Recognition accuracy is calculated using correctly identified faces divided by available faces in the class, multiplied by 100.Available faces are students present and detectable during attendance, rather than all students enrolled in the dataset.
- Evaluation: The analysis defines total faces, detected faces, available-in-training students, correctly identified students, and accuracy for individual test images.Accuracy is defined as correctly identified faces out of faces available in training.
VI. IMPLEMENTATION DETAILS
The implementation uses Google Colab and separates research files into folders for training, testing, processing, and prediction.
- VI. IMPLEMENTATION DETAILS: Google Colab hosts the research tasks, with separate folders organizing training, testing, processing, and prediction files.The training folder stores student images, while testing contains classroom images used for predictions.
VII. DISCUSSIONS
ClassVision reports classroom face detection and recognition without requiring students to be close to the camera, while identifying image-quality and dataset-bias limitations.
- VII. DISCUSSIONS: Incorrect predictions were associated with obstructions, partially visible faces, and students sitting too far away, especially in large classrooms.Examples of obstructions include laptops, bags, and bottles.
- VII. DISCUSSIONS: The training dataset may be biased by caps, facial hair, thick beards, and limited left- and right-view facial images.The paper states that these factors could lead to incorrect predictions.
- VII. DISCUSSIONS: Future enhancements include higher-resolution data capture, expanded appearance variations, new recognition models, LMS integration, and multimodal biometrics.The proposed multimodal extensions include voice or gesture recognition.
- VIII. CONCLUSION: ClassVision detects and recognizes classroom faces without requiring individuals to stand close to the camera.The conclusion describes preprocessing, RetinaFace embeddings, and 50x50-pixel ROI resizing as part of this approach.