Siamese Networks for Image Comparison and Discrepancy Localization
Pinak Paliwal · 2021 4th International Conference on Artificial Intelligence and Pattern Recognition · 2021
With the prevalence of online learning and exams, a side camera view overlooking a student's computer screen and work area is recommended to compare the expected view on exam screen to that of student's screen and identify cheating. This work proposes a novel architecture of Siamese convolutional neural networks to automate the proctoring process by simultaneously detecting disallowed activities on computer screen and localizing the region in cases where such activities are carried out in a portion of screen only. While the proposed Siamese detector compares the base image to test image to detect features in the entire image to ascertain discrepancy, the Siamese localizer splits the image area in sections to perform regional feature comparisons to yield discrepant regions. Both components run in a single-shot mode and avoid multi-shot regressions. On a public web screenshot database, the proposed Siamese detector and localizer both achieve impressive results with detector achieving 96.4% accuracy on 20-way full-screen discrepancy detection task and localizer achieving 94.4% accuracy along with mean overall IoU of 0.72 over discrepant segments.