Implementation of LiteRT-Based Waste Classification on Android Using Teachable Machine

Muhammad Yusuf, Haidar Khairullah (2026) Implementation of LiteRT-Based Waste Classification on Android Using Teachable Machine. Project Report. [s.n]. (Unpublished)

Text 05111942000022-Project_Report.pdf Restricted to Repository staff only Download (3MB)

Abstract

Effective municipal solid waste management is fundamental to environmental sustainability and public health. While source-side waste segregation is a critical stage in optimizing recycling pipelines, its practical implementation still relies heavily on manual sorting methods that are inherently inefficient, slow, and prone to high cross-contamination rates. To address these operational bottlenecks, this research designs, deploys, and evaluates a client-side Edge Artificial Intelligence (Edge AI) framework integrated into a native Android application. The computer vision architecture utilizes a MobileNetV2 convolutional backbone fine-tuned through transfer learning and regularized with a conservative learning rate of 5 × 10⁻⁵ to minimize localized dataset overfitting. The regularized model is compiled and executed entirely on-device using Google's high-performance LiteRT inference engine, ensuring low-latency, cloud-independent operation. Beyond the system implementation, the localized Edge AI module was subjected to a rigorous empirical evaluation consisting of 162 controlled test iterations to systematically assess classification performance under varying environmental conditions and operational constraints. The system achieved a baseline classification accuracy of 51.85%. Spatial evaluation revealed that performance was strongly influenced by focal distance, with the highest accuracy of 62.22% obtained at a close range of 15 cm due to target object dominance within the image frame, decreasing to 40.00% at a distance of 60 cm where dominant background features caused global feature degradation during the average pooling stage. Furthermore, illumination testing demonstrated a direct relationship between lighting conditions and feature extraction stability, while comparative analysis of runtime execution threads showed that the Asynchronous Static Image Pipeline eliminated user interface flickering by filtering temporal variations caused by involuntary hand micro-tremors. Overall, these findings demonstrate that decentralized, client-side Edge AI deployment on consumer smartphones provides a computationally lightweight, practical, and highly scalable solution for automated household waste classification.
=================================================================================================================================
Pengelolaan sampah padat perkotaan yang efektif merupakan aspek mendasar dalam mendukung keberlanjutan lingkungan dan kesehatan masyarakat. Meskipun pemilahan sampah dari sumber merupakan tahap penting untuk mengoptimalkan proses daur ulang, implementasinya masih sangat bergantung pada metode pemilahan manual yang secara inheren tidak efisien, lambat, dan rentan terhadap tingginya tingkat kontaminasi silang. Untuk mengatasi kendala operasional tersebut, penelitian ini merancang, mengimplementasikan, dan mengevaluasi kerangka kerja Kecerdasan Buatan Tepi (*Edge Artificial Intelligence* atau *Edge AI*) berbasis klien yang terintegrasi ke dalam aplikasi Android native. Arsitektur visi komputer menggunakan backbone konvolusional MobileNetV2 yang disempurnakan melalui *transfer learning* dan diregularisasi dengan *learning rate* konservatif sebesar 5 × 10⁻⁵ untuk meminimalkan *overfitting* pada dataset lokal. Model hasil regularisasi dikompilasi dan dijalankan sepenuhnya pada perangkat menggunakan mesin inferensi LiteRT berkinerja tinggi dari Google sehingga mampu beroperasi tanpa ketergantungan pada komputasi *cloud* dan dengan latensi yang sangat rendah. Selain implementasi sistem, modul *Edge AI* yang dikembangkan diuji melalui 162 iterasi pengujian terkontrol untuk mengevaluasi secara sistematis kinerja klasifikasi pada berbagai kondisi lingkungan dan kendala operasional. Sistem memperoleh akurasi klasifikasi dasar sebesar 51,85%. Evaluasi spasial menunjukkan bahwa kinerja sangat dipengaruhi oleh jarak fokus, dengan akurasi tertinggi sebesar 62,22% pada jarak 15 cm karena objek target mendominasi bidang citra, kemudian menurun menjadi 40,00% pada jarak 60 cm akibat dominasi latar belakang yang menyebabkan degradasi fitur global pada lapisan *average pooling*. Selain itu, pengujian pencahayaan menunjukkan adanya hubungan langsung antara tingkat luminansi dan stabilitas ekstraksi fitur, sedangkan analisis komparatif terhadap *runtime execution thread* membuktikan bahwa *Asynchronous Static Image Pipeline* mampu menghilangkan kedipan antarmuka visual dengan menyaring variasi temporal yang disebabkan oleh getaran mikro pada tangan. Secara keseluruhan, hasil penelitian ini menunjukkan bahwa penerapan *Edge AI* berbasis klien pada telepon pintar memberikan solusi yang ringan secara komputasi, layak diterapkan, dan sangat skalabel untuk klasifikasi sampah rumah tangga secara otomatis.

Item Type:	Monograph (Project Report)
Uncontrolled Keywords:	Edge AI, LiteRT, MobileNetV2, Waste Classification, Computer Vision, Android Deployment, Klasifikasi Sampah, Implementasi Android
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1637 Image processing--Digital techniques. Image analysis--Data processing.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Informatics Engineering > 55201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Yusuf Haidar Khairullah
Date Deposited:	23 Jul 2026 08:53
Last Modified:	23 Jul 2026 08:53
URI:	http://repository.its.ac.id/id/eprint/137321

Actions (login required)

View Item