Electoral Symbols and Vote Recognition on Paper Ballots- A Case Study of Nepal's General Election

Abstract

In electoral processes, which primarily use paper-based voting methods, this paper explores the application of advanced object detection technologies to improve the accuracy and efficiency of voting. Specifically, the goal is to detect electoral symbols and votes on paper ballots. We utilized two advanced models, Faster R-CNN with ResNet50 with Feature Pyramid Network and YOLOv8n, to develop a comprehensive ballot paper detection system with three modules: ballot creation, symbol detection, and vote validation module. The model for object detection is trained using a dataset of voted paper ballots, ensuring its accuracy and reliability. Ballot creation generates the ballot papers that feed into the symbols detection module. Symbol detection uses the trained model that recognizes and predicts symbols and votes on unseen ballots, a crucial step to see which was more effective at recognizing election symbols and accurately counting votes. The vote validation module checks the validity of votes and then calculates the overall number of votes obtained by each party symbol. Before running through the system, voted ballots are scanned and saved as image files in the ballot paper detection system. A ballot creation system that generates uniform ballot papers is necessary to preserve consistency in the voting process and ensure correct vote validation through a grid-cell design on the ballots.

Our comparison with the two models shows that Faster R-CNN with ResNet50 with FPN outperforms YOLOv8n, with a mean average precision of 93% over 92% for YOLOv8n. Despite its slower processing speed, Faster R-CNN's precision in recognizing votes is ideal for cases requiring high accuracy. This capacity can be useful in an electoral system that uses ballot paper for voting, where real-time processing demand is not required. This work shows the effectiveness of incorporating object detection technology into electoral systems and emphasizes the potential for larger applications to promote more transparent and fair elections.