Pengembangan Model Enhanced Faster R-CNN pada Tahap Region of Interest (RoI) Pooling untuk Otomatisasi Deteksi Cacat Baja pada Skala Industri Kecil Menengah (IKM)

Salsabilla, Nasywa (2026) Pengembangan Model Enhanced Faster R-CNN pada Tahap Region of Interest (RoI) Pooling untuk Otomatisasi Deteksi Cacat Baja pada Skala Industri Kecil Menengah (IKM). Other thesis, Institut Teknologi Sepuluh Nopember.

[thumbnail of 5010221188-Undergraduate_Thesis.pdf] Text
5010221188-Undergraduate_Thesis.pdf - Accepted Version
Restricted to Repository staff only

Download (6MB) | Request a copy

Abstract

Industri Kecil dan Menengah (IKM) logam di Desa Ngingas, di bawah naungan Koperasi Waru Buana Putra, menghadapi tantangan dalam menjaga konsistensi kualitas produk akibat ketergantungan pada inspeksi visual manual. Dengan volume produksi yang mencapai 2.000 unit per hari, keterbatasan manusia dalam proses inspeksi berpotensi menyebabkan terjadinya missed detection. Untuk mengatasi hal tersebut, penelitian ini mengembangkan pendekatan automated defect detection berbasis Faster R-CNN untuk mendukung proses inspeksi cacat pada permukaan baja. Penelitian ini bertujuan untuk mengevaluasi performa model Faster R-CNN dengan variasi metode Region of Interest (RoI) dan ukuran bin size dalam mendeteksi cacat permukaan baja. Metode yang diuji meliputi RoI Pooling, RoI Align, dan Weighted RoI Pooling dengan variasi bin size tertentu. Evaluasi dilakukan menggunakan detection rate, false detection rate (FDR), F1-score, serta frame per second (FPS) untuk menilai keseimbangan antara akurasi deteksi dan kecepatan inferensi. Hasil penelitian menunjukkan bahwa performa model masih dipengaruhi oleh keterbatasan jumlah dataset, ketidakseimbangan kelas, dan karakteristik cacat berukuran kecil. Berdasarkan evaluasi multi-kriteria, konfigurasi RoI Pooling dengan bin size 7 memperoleh skor tertinggi di antara konfigurasi yang diuji, dengan detection rate sebesar 27,17%, FDR sebesar 87,98%, dan kecepatan inferensi sebesar 38,39 FPS. Hasil tersebut menunjukkan bahwa konfigurasi ini memiliki keseimbangan relatif terbaik dalam ruang eksperimen, terutama dari sisi kecepatan inferensi. Namun, rendahnya detection rate dan tingginya FDR menunjukkan bahwa model belum memadai untuk diterapkan sebagai sistem inspeksi otomatis secara penuh. Oleh karena itu, konfigurasi tersebut diposisikan sebagai baseline untuk optimasi lanjutan dan uji coba terkontrol serta sebagai dasar awal pengembangan sistem pencatatan kualitas digital pada IKM.
=================================================================================================================================
Industrial Small and Medium Enterprises (SMEs) in the metal sector in Ngingas Village, under the coordination of the Waru Buana Putra Cooperative, face significant challenges in maintaining consistent product quality due to reliance on manual visual inspection. With a production volume reaching up to 2,000 units per day, human limitations in inspection processes may lead to missed defect detection. To address this issue, this study develops an automated defect detection approach based on Faster R-CNN to support defect inspection on steel surface products. This study aims to evaluate the performance of the Faster R-CNN model with variations of Region of Interest (RoI) methods and bin size configurations in detecting steel surface defects. The methods evaluated include RoI Pooling, RoI Align, and Weighted RoI Pooling with different bin size settings. Model performance is assessed using detection rate, false detection rate (FDR), F1-score, and frames per second (FPS) to evaluate the balance between detection accuracy and inference speed. The results indicate that model performance is still constrained by the limited dataset size, class imbalance, and the small-scale characteristics of the defects. Based on the multi-criteria evaluation, the RoI Pooling configuration with a bin size of 7 achieved the highest score among the tested configurations, with a detection rate of 27.17%, an FDR of 87.98%, and an inference speed of 38.39 FPS. These results show that this configuration provides the best relative balance within the experimental scope, particularly in terms of inference speed. However, the low detection rate and high FDR indicate that the model is not yet adequate for full deployment as an automated inspection system. Therefore, this configuration is positioned as a baseline for further optimization and controlled testing, as well as an initial foundation for developing a digital quality-recording system for SMEs.

Item Type: Thesis (Other)
Subjects: Q Science > Q Science (General) > Q325.5 Machine learning. Support vector machines.
Divisions: Faculty of Industrial Technology > Industrial Engineering > 26201-(S1) Undergraduate Thesis
Depositing User: Nasywa Salsabilla
Date Deposited: 31 Jul 2026 01:10
Last Modified: 31 Jul 2026 01:10
URI: http://repository.its.ac.id/id/eprint/140078

Actions (login required)

View Item View Item