طارق مصطاف رضوان
ABSTRACT
Cheating in the exam is defined as the desire to obtain high grades that are not worthy of the cheating student. Almost all educational programs evaluate the understanding of the materials by the class takers through in-class or online tests. Usually, these tests require monitoring for the entire period of the allowed time. Manual abnormal behavior detection methods may no longer be completely effective in fully preventing fraud during exams. In particular, abnormal behavior involves a wide range of activities such as passing the paper, peering through others’ papers, and signaling to others. The motivations of this research are to reduce cheating during exams, discover different types of cheating, and reduce the cost of monitoring on exams. The contributions of this work are the application of an intelligent approach using deep learning to effectively detect suspicious behaviors in the exam in real-time, and theََimplemented algorithm detects any suspicious behavior in the exams based on a learned object detection agent on a custom build image dataset that collects all possible cheating attempts in a classroom. A dataset named Tariquot Set was created and developed, which consisted of 12800 images taken in different environments in terms of environmental conditions, lighting and number of students. The dataset was divided into 70% for training and 30% for testing. During the Computer Vision Annotation Tool (CVAT) and the detection stage through the You Only Look Once version4 (YOLO v4) algorithm, the cheating sheets were divided into nine classes, which are (signals, talk, turning, standing, looking at another student’s answer, exchanging things, using the mobile, using the rivet, and opening or raising the answer sheet for another student to look at). Include experimental results that obtained from model inference which shows high cheating detection accuracy which approximately equals to 97.81%, and this result proves our benchmarks results and shows that our platform has higher accuracy than that of alternative models.
