Perancangan Realtime Vision System Berbasis Faster Region-Based Convolutional Neural Network (Faster RCNN) Untuk Klasifikasi Tipe Resin Sampah Plastik

Marsuki, Muhammad Faturrahman (2026) Perancangan Realtime Vision System Berbasis Faster Region-Based Convolutional Neural Network (Faster RCNN) Untuk Klasifikasi Tipe Resin Sampah Plastik. Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5022221058-Undergraduate_Thesis.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Peningkatan volume limbah plastik menuntut otomatisasi penyortiran yang cepat dan presisi. Penelitian ini mengusulkan perancangan real-time vision system berbasis Improved Faster R-CNN (backbone ResNet-50-vd yang dimodifikasi dengan DCNv2, CoordConv, dan SPP) untuk mengklasifikasikan 7 tipe resin plastik dan 1 kelas non-plastik. Ketidakseimbangan data latih diatasi melalui augmentasi stokastik yang berhasil menekan Mean Absolute Deviation (MAD) secara proporsional menjadi 1,35%. Untuk mencapai kecepatan operasional waktu nyata (>30 FPS), model dioptimasi menggunakan Taylor Pruning 25% dan Kuantisasi INT8. Kinerja sistem kemudian divalidasi secara end-to-end dalam lingkungan Co-Simulation menggunakan MuJoCo dan Simulink. Hasil pengujian menunjukkan arsitektur model Improved Baseline mampu mencapai tingkat Recall 99% dan Mean Average Precision (mAP) sebesar 90% dengan kecepatan awal 6 FPS. Implementasi kompresi ganda (Pruning dan Kuantisasi INT8) terbukti sukses mempertahankan mAP di angka 90% tanpa memicu degradasi akurasi lanjutan, serta sukses menyusutkan kebutuhan memori secara masif dari 237,9 MB menjadi 49,2 MB. Kompresi ini turut melesatkan kecepatan inferensi hingga mencapai 75 FPS. Pengujian Co-Simulation mengonfirmasi bahwa sistem kendali open-loop berhasil menerjemahkan bounding box menjadi aktuasi pneumatik (air jet) secara dinamis, membuktikan kelayakan implementasi arsitektur ini untuk otomasi daur ulang secara real-time.
======================================================================================================================================
The increasing volume of plastic waste demands fast and precise automated sorting. This research proposes the design of a real-time vision system based on Improved Faster R-CNN (ResNet-50-vd backbone modified with DCNv2, CoordConv, and SPP) to classify 7 types of plastic resins and 1 non-plastic class. Training data imbalance was addressed through stochastic augmentation, successfully reducing the Mean Absolute Deviation (MAD) to 1.35%. To achieve real-time operational speed (>30 FPS), the model was optimized using 25% Taylor Pruning and INT8 Quantization. The system's performance was then validated end-to-end in a Co-Simulation environment using MuJoCo and Simulink. Test results showed the Improved Baseline architecture achieved a 99% Recall and a Mean Average Precision (mAP) of 90% at an initial speed of 6 FPS. The implementation of dual compression (Pruning and INT8 Quantization) successfully maintained the mAP at 90% without further accuracy degradation, massively shrinking the memory footprint from 237.9 MB to 49.2 MB. This compression also boosted the inference speed up to 75 FPS. Co Simulation testing confirmed that the open-loop control system successfully translated bounding boxes into dynamic pneumatic (air jet) actuation, proving the feasibility of implementing this architecture for real-time recycling automation.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	Pemilahan Plastik, Faster R-CNN, Co-Simulation, Kuantisasi, Pruning Plastic Waste. Faster R-CNN, Co-Simulation, Quantization, Pruning
Subjects:	T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK3070 Automatic control T Technology > TK Electrical engineering. Electronics Nuclear engineering > TK7882.P3 Pattern recognition systems
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Electrical Engineering > 20201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Faturrahman Marsuki
Date Deposited:	27 Jul 2026 03:29
Last Modified:	27 Jul 2026 03:29
URI:	http://repository.its.ac.id/id/eprint/137799

Actions (login required)

View Item