Implementasi YOLOv5m dengan Transfer Learning dan Modifikasi Fine Tuning Berakurasi Tinggi untuk Deteksi Kematangan Fragaria x ananassa Tristar

Saepul Rohman, - and Roer Eka Pawinanto, - (2025) Implementasi YOLOv5m dengan Transfer Learning dan Modifikasi Fine Tuning Berakurasi Tinggi untuk Deteksi Kematangan Fragaria x ananassa Tristar. S1 thesis, Universitas Pendidikan Indonesia.

Abstract

Fragaria x ananassa atau stroberi merupakan komoditas hortikultura dengan potensi ekonomi tinggi. Namun, proses panen membutuhkan ketepatan waktu dan biaya tenaga kerja yang mencapai 73-80% dari total biaya produksi menjadi tantangan utama bagi petani. Kekurangan teknologi dalam proses menentukan kematangan disebabkan kondisi pencahayaan yang bervariasi, serta kurangnya dataset buah publik yang mengakibatkan akurasi rendah, kecepatan lambat. Robot pemanen yang ditunjang dengan kemampuan algortima visi komputer yang ringan, akurat dan efisien dalam menentukan tingkat kematangan stroberi menjadi solusi penting untuk mengatasi kekurangan tenaga kerja dan memerlukan algoritma penglihatan mesin yang ringan dan efisien. Modifikasi algoritma visi komputer untuk memprediksi kematangan stroberi selama ini masih memiliki beberapa kekurangan yaitu generalisasi model, deteksi gagal, kesulitan dalam mendeteksi buah yang tersembunyi atau tumpang tindih. Oleh karena itu, penelitian ini mengembangkan sistem deteksi kematangan buah stroberi tristar menggunakan teknologi deep learning berbasis model YOLOv5m yang dioptimalkan melalui pendekatan transfer learning dan fine-tuning untuk mengatasi kekurangan algoritma visi komputer pada riset-riset sebelumnya. Model YOLOv5m dimodifikasi menggunakan transfer learning dengan pretrained weights dan dilakukan fine-tuning dalam dua tahap: 50 epochs dengan freeze 10 lapisan pertama, dilanjutkan 200 epochs dengan unfreeze seluruh lapisan. Hasil evaluasi menunjukkan model mencapai precision 94,3%, recall 92,8%, F1-score 93%, dan mean average precision (mAP) 97,5% dengan kecepatan deteksi 75,30 FPS. Sistem aplikasi android berhasil deteksi real-time dengan respons rata-rata 3 detik. Hasil evaluasi dalam penelitian ini melampaui hasil evaluasi riset-riset sebelumnya sehingga dapat dijadikan referensi. Selain itu, Penelitian ini memberikan kontribusi dalam pengembangan sistem pertanian cerdas sehingga berpotensi diintegrasikan dalam robot pemanen otomatis. Fragaria x ananassa or strawberry is a horticultural commodity with high economic potential. However, the harvesting process requires timeliness and labor costs that reach 73-80% of the total production costs are the main challenges for farmers. Lack of technology in the process of determining ripeness is due to varying lighting conditions, as well as the lack of public fruit datasets resulting in low accuracy, slow speed. Harvesting robots supported by lightweight, accurate and efficient computer vision algorithm capabilities in determining strawberry ripeness are an important solution to overcome labor shortages and require lightweight and efficient machine vision algorithms. Modifications to computer vision algorithms to predict strawberry ripeness so far still have several shortcomings, namely model generalization, failed detection, difficulty in detecting hidden or overlapping fruits. Therefore, this study developed a ripeness detection system for tristar strawberries using deep learning technology based on the YOLOv5m model which was optimized through a transfer learning and fine-tuning approach to overcome the shortcomings of computer vision algorithms in previous studies. The YOLOv5m model was modified using transfer learning with pretrained weights and fine-tuning was carried out in two stages: 50 epochs with freezing the first 10 layers, followed by 200 epochs with unfreezing all layers. The evaluation results showed that the model achieved a precision of 94.3%, a recall of 92.8%, an F1-score of 93%, and a mean average precision (mAP) of 97.5% with a detection speed of 75.30 FPS. The android application system successfully detected in real time with an average response of 3 seconds. The evaluation results in this study exceeded the evaluation results of previous studies so that they can be used as a reference. In addition, this study contributes to the development of intelligent agricultural systems so that they have the potential to be integrated into automatic harvesting robots.

	Text S_PTOIR_2103754_Title.pdf Download (595kB)
	Text S_PTOIR_2103754_Chapter1.pdf Restricted to Staf Perpustakaan Download (297kB) | Request a copy
	Text S_PTOIR_2103754_Chapter2.pdf Restricted to Staf Perpustakaan Download (340kB) | Request a copy
	Text S_PTOIR_2103754_Chapter3.pdf Restricted to Staf Perpustakaan Download (840kB) | Request a copy
	Text S_PTOIR_2103754_Chapter4.pdf Restricted to Staf Perpustakaan Download (1MB) | Request a copy
	Text S_PTOIR_2103754_Chapter5.pdf Download (231kB)
	Text S_PTOIR_2103754_Appendix.pdf Restricted to Staf Perpustakaan Download (790kB) | Request a copy

Item Type:	Thesis (S1)
Additional Information:	https://scholar.google.com/citations?hl=id&user=Je6VaaEAAAAJ ID SINTA Dosen Pembimbing Roer Eka Pawinanto: 6745886
Uncontrolled Keywords:	deep learning, YOLOv5, deteksi kematangan buah, stroberi tristar, transfer learning, fine-tuning, aplikasi android deep learning, YOLOv5, Fruit Ripeness Detection, Tristar Strawberry, Transfer Learning, Fine-tuning, Android Application
Subjects:	L Education > L Education (General) T Technology > TK Electrical engineering. Electronics Nuclear engineering
Divisions:	Fakultas Pendidikan Teknik dan Industri > S1 Pendidikan Teknik Otomasi Industri dan Robotika
Depositing User:	Saepul Rohman
Date Deposited:	24 Oct 2025 05:31
Last Modified:	24 Oct 2025 05:31
URI:	http://repository.upi.edu/id/eprint/143695

Actions (login required)

View Item