Pengembangan Sistem Quality Inspection pada Cacat Visual Produk Beras Berbasis Computer Vision untuk Mengidentifikasi Defect Rate

Filardi, Achmad Fakhri Filardi (2026) Pengembangan Sistem Quality Inspection pada Cacat Visual Produk Beras Berbasis Computer Vision untuk Mengidentifikasi Defect Rate. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5010221181-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Beras merupakan komoditas pangan strategis dengan konsumsi nasional 111,32 kg per kapita per tahun. Seiring meningkatnya pendapatan masyarakat, preferensi konsumen bergeser ke beras premium bermutu visual lebih ketat, sehingga industri dituntut menjaga konsistensi kualitas beras. Namun, inspeksi mutu beras masih banyak dilakukan secara manual sehingga rentan terhadap human error. Penelitian ini bertujuan mengembangkan sistem quality inspection berbasis computer vision untuk mendeteksi cacat visual dan menghitung defect rate beras secara otomatis, mengacu SNI 6128:2020 dengan empat kelas: beras normal, berubah warna, tidak sesuai ukuran, dan kotoran. Penelitian menggunakan pendekatan Research and Development dengan model YOLOv8n-seg (instance segmentation) untuk deteksi objek dan library OpenCV untuk pre-processing citra. Data dikumpulkan menggunakan kamera smartphone 4K dan dianotasi melalui platform Roboflow dengan pendekatan transfer learning. Evaluasi dilakukan dalam tiga tahap: per butir menggunakan Confusion Matrix (Accuracy, Precision, Recall, F1-Score), per batch menggunakan Mean Absolute Error (MAE), dan keandalan sistem pengukuran menggunakan Gauge Repeatability and Reproducibility (Gauge R&R) dengan pendekatan Average and Range berdasarkan standar AIAG. Pengujian dilakukan pada 5 batch sampel (100 gram per batch) dengan tingkat defect rate acuan 5%, 10%, 15%, 20%, dan 25%. Hasil evaluasi menunjukkan performa terbaik pada kelas Beras Normal (precision 52,55%, recall 81,52%, F1-Score 63,93%), sedangkan rata-rata makro seluruh kelas hanya 32,77%/26,16%/24,83% akibat performa rendah pada kelas cacat. MAE sebesar 47,38% dan %GRR sebesar 90,16% (Not Acceptable) mengindikasikan domain gap antara data pelatihan dan data pengujian sebagai akar permasalahan utama, sehingga sistem belum dapat diandalkan untuk menggantikan inspeksi manual.
=====================================================================================================================================
Rice is a strategically important food commodity in Indonesia, with national consumption reaching 111.32 kg per capita per year. As household income rises, consumer preference is shifting toward premium rice with stricter visual quality standards, requiring the food industry to consistently maintain rice quality. However, rice quality inspection is still largely performed manually, making it prone to human error. This study aims to develop a computer vision-based quality inspection system to automatically detect visual defects and calculate the defect rate in rice, referring to the SNI 6128:2020 standard across four classes: normal, discolored, off-size, and impure rice. The study follows a Research and Development approach using the YOLOv8n-seg (instance segmentation) model for object detection and the OpenCV library for image pre-processing. Data were collected using a 4K smartphone camera and annotated on the Roboflow platform using a transfer learning approach. Evaluation was conducted in three stages: per-grain evaluation using a Confusion Matrix (Accuracy, Precision, Recall, F1-Score), per-batch evaluation using Mean Absolute Error (MAE), and measurement system reliability evaluation using Gauge Repeatability and Reproducibility (Gauge R&R) with the Average and Range method based on the AIAG standard. Testing was conducted on 5 sample batches (100 grams each) with reference defect rates of 5%, 10%, 15%, 20%, and 25%. The results show the best performance on the Normal Rice class (precision 52.55%, recall 81.52%, F1-Score 63.93%), while the macro-average across all classes was only 32.77%/26.16%/24.83% due to lower performance on the defect classes. An MAE of 47.38% and a %GRR of 90.16% (Not Acceptable) indicate a domain gap between training and testing data as the primary root cause, meaning the system is not yet reliable enough to replace manual inspection.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Beras, Computer Vision, Confusion Matrix, Deep Learning, Defect Rate, Gauge R&R, Inspeksi Kualitas, Mean Absolute Error, YOLOv8, Computer Vision, Confusion Matrix, Deep Learning, Defect Rate, Mean Absolute Error, Quality Inspection, Rice, YOLOv8
Subjects:	H Social Sciences > HD Industries. Land use. Labor > HD3656 Inspection. Factory inspection H Social Sciences > HD Industries. Land use. Labor > HD62.15 Total quality management. T Technology > TS Manufactures > TS156 Quality Control. QFD. Taguchi methods (Quality control)
Divisions:	Faculty of Industrial Technology and Systems Engineering (INDSYS) > Industrial Engineering > 26201-(S1) Undergraduate Thesis
Depositing User:	Achmad Fakhri Filardi
Date Deposited:	30 Jul 2026 08:44
Last Modified:	30 Jul 2026 08:44
URI:	http://repository.its.ac.id/id/eprint/139711

Actions (login required)

View Item