Perbandingan YOLOv8 dengan YOLOv11 pada Sistem Pendeteksi Kerusakan Jahitan Kantong Pupuk

Daniswara, Evan Firjatullah (2026) Perbandingan YOLOv8 dengan YOLOv11 pada Sistem Pendeteksi Kerusakan Jahitan Kantong Pupuk. Other thesis, Institut Tenologi Sepuluh Nopember.

Text 2040221053-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (6MB) | Request a copy

Abstract

PT Petrokimia Gresik (PKG) merupakan salah satu produsen pupuk terbesar di Indonesia dengan kapasitas produksi mencapai 8,9 juta ton per tahun. Kualitas jahitan karung menjadi faktor penting untuk mencegah kebocoran, kontaminasi, dan kerugian material pada proses pengantongan. Kerusakan jahitan umumnya berukuran relatif kecil, rata-rata sekitar 7,3% dari luas citra, dan memiliki kemiripan visual dengan tekstur karung, sehingga kondisi ini menyulitkan proses deteksi pada operasional nyata. Proyek Akhir ini merancang sistem deteksi kerusakan jahitan menggunakan pendekatan deteksi bertingkat (cascade). Model 1 untuk mendeteksi lokasi karung dan model 2 untuk mendeteksi kondisi jahitan pada citra hasil crop serta membandingkan performa YOLOv8 dan YOLOv11 pada kedua model berdasarkan precision, recall, mAP50, mAP50-95. Sistem terintegrasi dengan kamera IP, NodeMCU ESP8266, dan protokol MQTT untuk memberikan peringatan melalui tower lamp dan buzzer. Hasil pengujian model 1 menunjukkan YOLOv8 mencapai mAP50 0,9915 dan YOLOv11 mencapai mAP50 0,9916, sedangkan pada model 2, YOLOv8 mencapai mAP50 0,957 dan YOLOv11 mencapai mAP50 0,959, dengan mAP50-95 YOLOv11 yang juga lebih tinggi pada kedua model yaitu model 1 0,8494 berbanding 0,8481 selisih 0,0013 dan model 2 0,577 berbanding 0,574 selisih 0,003, begitu pula pada recall model 1 0,9600 berbanding 0,9498 selisih 0,0102, model 2 0,938 berbanding 0,922 selisih 0,016. Sebagai pembanding, SSD-MobileNet V1 dan V2 hanya mencapai mAP 0,62–0,64, jauh di bawah performa YOLO. mAP50 kedua kelas jahitan berada pada rentang 0,947–0,971 pada kedua arsitektur. Secara keseluruhan, YOLOv11 unggul pada recall dan mAP50-95 dibandingkan YOLOv8 pada kedua model, sementara YOLOv8 unggul pada precision yaitu model 1 0,9905 berbanding 0,9850 selisih 0,0055, model 2 0,926 berbanding 0,909 selisih 0,017. Mengingat selisih performa kedua arsitektur pada seluruh metrik berada di bawah 2% yaitu 0,0001–0,017, YOLOv8 tetap menjadi alternatif yang layak dipertimbangkan, terutama untuk pengujian lebih lanjut pada sistem cascade secara menyeluruh.
==================================================================================================================================
PT Petrokimia Gresik (PKG) is one of the largest fertilizer producers in Indonesia, with a production capacity of 8.9 million tons per year. Sack stitching quality is a critical factor in preventing leakage, contamination, and material loss during the bagging process. Stitching defects are typically small in size, averaging around 7.3% of the image area, and are visually similar to the surrounding sack texture, making detection more challenging under real operational conditions. This Final Project designs a stitching defect detection system using a cascade approach. Model 1 detects the location of the sack, and model 2 detects the stitching condition on the cropped image, and compares the performance of YOLOv8 and YOLOv11 on both models based on precision, recall, mAP50, and mAP50-95. The system is integrated with an IP camera, NodeMCU ESP8266, and the MQTT protocol to deliver alerts via a tower lamp and buzzer. Model 1 testing results show that YOLOv8 achieves an mAP50 of 0.9915 and YOLOv11 achieves an mAP50 of 0.9916, while for model 2, YOLOv8 achieves an mAP50 of 0.957 and YOLOv11 achieves an mAP50 of 0.959, with YOLOv11 also showing higher mAP50-95 on both models, namely Model 1 0.8494 versus 0.8481 a difference of 0.0013 and model 2 0.577 versus 0.574 a difference of 0.003, as well as higher recall, model 1 0.9600 versus 0.9498 a difference of 0.0102, model 2 0.938 versus 0.922 a difference of 0.016. As a comparison, SSD-MobileNet V1 and V2 only achieve an mAP of 0.62–0.64, far below the performance of YOLO. The mAP50 for both stitching classes ranges from 0.947 to 0.971 across both architectures. Overall, YOLOv11 outperforms YOLOv8 in recall and mAP50-95 on both models, while YOLOv8 outperforms YOLOv11 in precision, namely model 1 0.9905 versus 0.9850 a difference of 0.0055, model 2 0.926 versus 0.909 a difference of 0.017. Given that the performance gap between the two architectures across all metrics is below 2% that is 0.0001–0.017, YOLOv8 remains a viable alternative worth considering, particularly for further testing on the complete cascade system.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Deteksi objek, Jahitan karung, Computer vision, YOLOv8, YOLOv11, Quality control. ======================================================================================================================== Object detection, Sack stitching, Computer vision, YOLOv8, YOLOv11, Quality control
Subjects:	T Technology > T Technology (General) T Technology > T Technology (General) > T385 Visualization--Technique T Technology > T Technology (General) > T57.5 Data Processing
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Evan Firjatullah Daniswara
Date Deposited:	14 Aug 2026 06:20
Last Modified:	14 Aug 2026 06:20
URI:	http://repository.its.ac.id/id/eprint/144349

Actions (login required)

View Item