Electronic Theses and Dissertation

Proses penghitungan suara merupakan aspek krusial dalam pelaksanaan pemilu yang menjamin transparansi dan kepercayaan publik. Meskipun Komisi Pemilihan Umum (KPU) telah mengembangkan sistem Sistem Informasi Rekapitulasi (Sirekap) sebagai alat bantu digital, sistem ini masih menghadapi tantangan teknis, terutama dalam akurasi pengenalan angka pada formulir C hasil. Penelitian ini bertujuan untuk mengembangkan sistem Optical Character Recognition (OCR) berbasis YOLOv8 (You Only Look Once Version 8) dan Long Short-Term Memory (LSTM) dengan Connectionist Temporal Classification (CTC) Loss guna mendeteksi dan mengenali angka hasil perolehan suara secara akurat pada formulir Sirekap pemilu presiden 2024. Metodologi yang digunakan mencakup dua tahap utama: (1) deteksi lokasi angka menggunakan model YOLOv8n, dan (2) pengenalan angka hasil deteksi menggunakan model OCR LSTM + CTC. Berdasarkan hasil evaluasi, model YOLOv8n berhasil mendeteksi angka dengan akurasi rata-rata sebesar 98,5%, presisi rata-rata sebesar 100%, recall rata-rata sebesar 98,5%, F1-score rata-rata sebesar 99,3%, serta Intersection over Union (IoU) rata-rata sebesar 91,3% pada dataset pengujian. Sementara itu, model OCR berbasis LSTM + CTC mencapai Character Error Rate (CER) sebesar 0,0406 dan Word Error Rate (WER) sebesar 0,1119 dengan total prediksi benar sebanyak 365 dari 411 dataset pengujian. Hasil ini menunjukkan bahwa pendekatan yang diusulkan mampu meningkatkan akurasi pengenalan angka pada formulir hasil pemilu dan berpotensi menjadi alternatif yang mampu dalam meningkatkan keandalan sistem digitalisasi hasil suara.

Kata kunci: Sirekap, Pemilu, YOLOv8, OCR, LSTM, CTC Loss

The vote-counting process is a crucial aspect of elections that ensures transparency and public trust. Although the General Elections Commission (KPU) has developed Recapitulation Information System (Sirekap) as a digital support system, it still faces technical challenges, particularly in the accuracy of digit recognition on the C-form results. This study aims to develop an Optical Character Recognition (OCR) system based on YOLOv8 (You Only Look Once Version 8) and Long Short-Term Memory (LSTM) with Connectionist Temporal Classification (CTC) Loss to accurately detect and recognize vote count digits on the 2024 Presidential Election Sirekap forms. The methodology consists of two main stages: (1) digit location detection using the YOLOv8n model, and (2) recognition of the detected digits using the LSTM + CTC-based OCR model. Based on the evaluation results, the YOLOv8n model successfully detected numbers with an average accuracy of 98.5%, an average precision of 100%, an average recall of 98.5%, an average F1-score of 99.3%, and an average Intersection over Union (IoU) of 91.3% on the test dataset. Meanwhile, the LSTM + CTC-based OCR model achieved a Character Error Rate (CER) of 0.0406 and a Word Error Rate (WER) of 0.1119, with 365 correct predictions out of 411 test dataset. These results indicate that the proposed approach is capable of improving digit recognition accuracy on election result forms and holds potential as an alternative to enhance the reliability of vote-count digitalization systems. Keywords: Sirekap, Pemilu, YOLOv8, OCR, LSTM, CTC Loss