Sistem Deteksi Sampah Padat Berbasis Convolutional Neural Network

Dio, Dio Lavyarel Alif Setiawan (2026) Sistem Deteksi Sampah Padat Berbasis Convolutional Neural Network. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 2040221065-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (3MB) | Request a copy

Abstract

Pengelolaan limbah padat di lingkungan laboratorium keteknikan masih sangat bergantung pada pemilahan manual yang lambat, tidak efisien, dan berisiko menimbulkan cedera akibat benda tajam yang tersembunyi di tumpukan sampah campuran. Algoritma deteksi objek berbasis Convolutional Neural Network (CNN) konvensional, seperti YOLO, sering mengalami kesulitan dalam membedakan objek dengan kemiripan visual yang tinggi serta gagal mendeteksi objek yang saling bertumpuk. Meskipun algoritma berbasis Transformer, seperti DETR, mampu mengatasi permasalahan tersebut melalui mekanisme Attention, beban komputasinya yang sangat besar membuatnya tidak dapat dijalankan secara real-time pada perangkat Edge Computing dengan sumber daya terbatas. Penelitian ini mengusulkan arsitektur hibrida yang menggabungkan Efficient Model V2 (EMOv2) sebagai backbone berbasis Spanning Window Attention dengan Neck dan Instance Segmentation Head dari YOLO26. Integrasi ini bertujuan menangkap konteks global citra sebagaimana pada Transformer, namun tetap mempertahankan jumlah parameter yang ringan agar kompatibel dengan perangkat Jetson Nano. Model dilatih menggunakan dataset citra limbah padat yang terdiri atas baut, daun, kaca, kaleng, mur, dan paku dengan resolusi 640 × 640 piksel. Hasil pengujian menunjukkan bahwa model usulan hanya menggunakan 5,4 juta parameter, tetapi mampu mencapai nilai Box mAP@50 sebesar 0,852 dan Mask mAP@50 sebesar 0,846. Meskipun beban komputasi (GFLOPs) meningkat akibat operasi Attention, model usulan menunjukkan efisiensi memori yang lebih baik dibandingkan varian YOLO Small dengan 11 juta parameter yang berisiko mengalami *out of memory* pada Jetson Nano. Implementasi algoritma Object Tracking BoT-SORT dan *Temporal Smoothing* pada inferensi video secara real-time terbukti efektif menghilangkan kedipan label serta menstabilkan deteksi objek yang saling bertumpuk. Penelitian ini membuktikan bahwa integrasi mekanisme Attention pada model berkategori Nano mampu menghasilkan akurasi yang sebanding dengan model berukuran lebih besar, sehingga menjadi solusi yang layak untuk sistem pemilahan limbah otomatis berbasis perangkat Edge.
===================================================================================================================================
The management of solid waste in engineering laboratory environments still relies heavily on manual sorting, which is slow, inefficient, and poses injury risks due to sharp objects concealed within mixed waste piles. Conventional Convolutional Neural Network (CNN)-based object detection algorithms, such as YOLO, often struggle to distinguish visually similar objects and fail to detect overlapping objects. Although Transformer-based algorithms, such as DETR, can address these challenges through the Attention mechanism, their substantial computational requirements prevent real-time deployment on resource-constrained Edge Computing devices. This study proposes a hybrid architecture that integrates Efficient Model V2 (EMOv2) as a lightweight backbone based on Spanning Window Attention with the Neck and Instance Segmentation Head of YOLO26. This integration aims to capture global image context similar to Transformer architectures while maintaining a lightweight parameter size suitable for deployment on Jetson Nano hardware. The model was trained using a solid waste image dataset consisting of bolts, leaves, glass, cans, nuts, and nails with a resolution of 640 × 640 pixels. Experimental results show that the proposed model uses only 5.4 million parameters while achieving a Box mAP@50 of 0.852 and a Mask mAP@50 of 0.846. Although the computational cost (GFLOPs) increases due to the Attention mechanism, the proposed model demonstrates superior memory efficiency compared with YOLO Small variants containing 11 million parameters, which are prone to *out-of-memory* issues on the Jetson Nano. Furthermore, the integration of the BoT-SORT object tracking algorithm and *Temporal Smoothing* during real-time video inference effectively eliminates label flickering and stabilizes the detection of overlapping objects. These findings demonstrate that incorporating Attention mechanisms into Nano-scale models can achieve accuracy comparable to larger models, making the proposed approach a practical solution for automated waste-sorting systems on Edge devices.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Attention Mechanism, Edge Computing, EMOv2, Instance Segmentation, Jetson Nano, YOLO26
Subjects:	Q Science Q Science > Q Science (General) > Q325.78 Back propagation Q Science > QA Mathematics > QA336 Artificial Intelligence Q Science > QA Mathematics > QA76.87 Neural networks (Computer Science) Q Science > QA Mathematics > QA76.F56 Data structures (Computer science) Q Science > QA Mathematics > QA9.58 Algorithms
Divisions:	Faculty of Vocational > 36304-Automation Electronic Engineering
Depositing User:	Dio Lavyarel Alif Setiawan
Date Deposited:	07 Aug 2026 04:04
Last Modified:	07 Aug 2026 04:04
URI:	http://repository.its.ac.id/id/eprint/144190

Actions (login required)

View Item