ANALISIS KOMPARATIF KINERJA MODEL RESNET50 DAN RESNET101 DALAM PENGEMBANGAN OCR MENGGUNAKAN DATASET CHARS74K

RIZAL FIKRI, . (2026) ANALISIS KOMPARATIF KINERJA MODEL RESNET50 DAN RESNET101 DALAM PENGEMBANGAN OCR MENGGUNAKAN DATASET CHARS74K. Sarjana thesis, UNIVERSITAS NEGERI JAKARTA.

	Text COVER.pdf Download (1MB)
	Text BAB I.pdf Download (254kB)
	Text BAB II.pdf Restricted to Registered users only Download (927kB) | Request a copy
	Text BAB III.pdf Restricted to Registered users only Download (911kB) | Request a copy
	Text BAB IV.pdf Restricted to Registered users only Download (1MB) | Request a copy
	Text BAB V.pdf Restricted to Registered users only Download (234kB) | Request a copy
	Text DAFTAR PUSTAKA.pdf Download (223kB)
	Text Lampiran Skripsi.pdf Restricted to Registered users only Download (2MB) | Request a copy

Abstract

Pengenalan karakter optik (OCR) merupakan komponen penting dalam digitalisasi dokumen, namun pemilihan arsitektur deep learning yang tepat masih menjadi tantangan akibat trade-off antara kompleksitas model dan kinerja aktual. Penelitian ini bertujuan membandingkan kinerja ResNet50 dan ResNet101 dalam proses pelatihan maupun hasil OCR pada dataset Chars74K. Metode yang digunakan meliputi preprocessing citra, pelatihan kedua arsitektur, serta pengujian statistik Wilcoxon Signed-Rank terhadap empat metrik pelatihan. Hasil menunjukkan ResNet50 unggul signifikan pada accuracy (88,8% vs 87,7%), loss, dan validation loss, serta lebih efisien dari segi waktu pelatihan (22,9% lebih cepat) dan inferensi (16,4% lebih cepat), sementara validation accuracy kedua model relatif setara. Namun, pengujian OCR dengan metrik WER dan CER menunjukkan hasil sebaliknya: ResNet101 mencapai tingkat kesalahan lebih rendah (WER 25,00%, CER 5,71%) dibandingkan ResNet50 (WER 37,50%, CER 11,43%). Kondisi ini diduga akibat ketidaksesuaian kapasitas model dengan kompleksitas dataset yang relatif sederhana. Dapat disimpulkan bahwa ResNet50 lebih efisien dan stabil dalam proses pelatihan, sedangkan ResNet101 menunjukkan keunggulan pada tahap inferensi OCR, sehingga pemilihan arsitektur perlu disesuaikan dengan prioritas antara efisiensi pelatihan dan akurasi hasil OCR. Kata kunci : Pengenalan Karakter Optik, Residual Network (ResNet), ResNet50, ResNet101, Chars74K, Word Error Rate, Character Error Rate ***** Optical Character Recognition (OCR) is an essential component in document digitization; however, selecting an appropriate deep learning architecture remains a challenge due to the trade-off between model complexity and actual performance. This study aims to compare the performance of ResNet50 and ResNet101 in both the training process and OCR results on the Chars74K dataset. The methods employed include image preprocessing, training of both architectures, and statistical testing using the Wilcoxon Signed-Rank test on four training metrics. The results show that ResNet50 significantly outperformed ResNet101 in terms of accuracy (88.8% vs 87.7%), loss, and validation loss, and was more efficient in training time (22.9% faster) and inference time (16.4% faster), while validation accuracy between the two models was relatively equivalent. However, OCR testing using WER and CER metrics showed the opposite result: ResNet101 achieved lower error rates (WER 25.00%, CER 5.71%) compared to ResNet50 (WER 37.50%, CER 11.43%). This condition is presumed to result from a mismatch between model capacity and the relatively simple complexity of the dataset. It can be concluded that ResNet50 is more efficient and stable during the training process, whereas ResNet101 demonstrates superior performance at the OCR inference stage, indicating that architecture selection should be adjusted according to the priority between training efficiency and OCR accuracy results. Keywords: Optical Character Recognition, Residual Network (ResNet), ResNet50, ResNet101, Chars74K, Word Error Rate, Character Error Rate

Item Type:	Thesis (Sarjana)
Additional Information:	1). Dr. Widodo, S.Kom., M.Kom. ; 2). Muchammad Ficky Duskarnaen, ST., M.Sc.
Subjects:	Teknologi dan Ilmu Terapan > Teknik Komputer
Divisions:	FT > S1 Pendidikan Teknik Informatika Komputer
Depositing User:	Rizal Fikri .
Date Deposited:	12 Aug 2026 04:02
Last Modified:	12 Aug 2026 04:02
URI:	http://repository.unj.ac.id/id/eprint/69973

Actions (login required)

View Item