AGENTS: Journal of Artificial Intelligence and Data Science

Design and Development of a Mobile Application for Indonesian Sign Language Translation Using YOLO

2026-09-14T06:02:29+00:00

Indonesian Sign Language (BISINDO) is one of the primary communication media for the deaf community in Indonesia. However, most people do not understand BISINDO, creating communication barriers. This study aims to design and develop a real-time BISINDO translator mobile application based on the You Only Look Once (YOLO) algorithm. The system is designed to recognize hand gestures representing letters and numbers in BISINDO through a smartphone camera. The research methodology includes BISINDO gesture image dataset collection, data preprocessing, YOLO model training, and model integration into an Android-based mobile application. The testing results show that the YOLO model can detect and classify BISINDO gestures with high accuracy and fast response time. The resulting mobile application can serve as a practical and efficient communication tool to bridge interactions between the deaf community and the general public.

EVALUATION OF PERMUTATION FEATURE IMPORTANCE METHOD FOR PLANT DISEASE DETECTION

2026-09-14T05:54:56+00:00

The development of machine learning models for plant disease detection often faces a trade-off between predictive accuracy and interpretability, where more accurate models tend to be more difficult to explain. This study evaluates that trade-off by comparing Random Forest and Neural Network models on the classification of potato and tomato leaf diseases using Permutation Feature Importance (PFI) as a model-agnostic interpretation instrument. The study used the PlantVillage dataset, consisting of 18,163 leaf images spread across 13 classes, with ten hand-crafted features covering color, texture, and leaf-area-ratio characteristics as the input representation for both models. The results show that the Neural Network achieved a test accuracy of 92.4% and a mean 5-fold cross-validation accuracy of 90.3%, outperforming the Random Forest, which achieved 82.2% and 82.3%, respectively. However, the PFI results reveal a low level of consistency in the most important features between the two models: Random Forest consistently relies on texture and color-area-ratio features that are biologically explainable, whereas the Neural Network relies more on global color intensity.

A Deep Learning-Based Visual Classification System for Pets and Pests Using Yolov8 in a Domestic Environment

2026-02-24T18:12:53+00:00

Lingkungan domestik merupakan area interaksi antara manusia dan berbagai jenis hewan, di mana kehadiran hama seperti tikus, kecoa, dan kadal dapat menyebabkan kerugian materi dan risiko kesehatan, termasuk leptospirosis. Hal ini relevan dengan prinsip perlindungan kehidupan (hifz an-nafs) dalam Islam, yang mendorong keselamatan manusia dari bahaya. Studi ini bertujuan untuk merancang dan menyalakan model deteksi objek cerdas berbasis pembelajaran mendalam yang dapat membedakan secara visual antara hewan peliharaan—seperti kucing, anjing, dan kelinci—dan hama, secara real-time, serta mengintegrasikannya dengan layanan notifikasi Telegram Bot untuk pemantauan praktis. Studi ini menggunakan pendekatan kuantitatif eksperimental dengan metode CRISP-DM (Cross Industry Standard Process for Data Mining) dan arsitektur YOLOv8, dengan dataset yang dari 3.600 gambar Model terbaik yang menggunakan optimizer AdamW menghasilkan akurasi 97,3%, konsisten 96,45%, recall 94,73%, dan F1-Score 95,57%, menunjukkan stabilitas pada epoch 90. Analisis matriks menunjukkan tentang yang tinggi di semua kelas, menunjukkan bahwa model tersebut mampu membedakan hewan peliharaan dari hama secara efektif. Sistem integrasi yang memungkinkan deteksi waktu nyata, manajemen data statistik harian, dan pengiriman notifikasi otomatis ke Telegram, menjadikan model ini solusi cerdas untuk memantau dan mengendalikan lingkungan domestik yang aman dan efisien.

IMPLEMENTASI ALGORITMA YOLO DAN TRACKING UNTUK KLASIFIKASI KONDISI KESEHATAN AYAM PETELUR DI DESA BELAJEN KABUPATEN ENREKANG

2026-02-06T14:08:26+00:00

his research develops a detection and classification system for the health conditions of laying hens
using the YOLOv8 algorithm combined with the ByteTrack tracking method. The system is designed
to identify three health categories—healthy, sick, and dead—based on image analysis from CCTV
footage and mobile phone recordings streamed via OBS as an intermediary. The dataset consists of
labeled images across these three categories, which were trained to produce the optimal detection
model.Testing was conducted offline using video data excluded from the training set. System
evaluation utilized precision, recall, and mean Average Precision (mAP) metrics. The results
demonstrate that the model achieves accurate detection and classification, specifically reaching an
mAP50 of 0.87 for dead chickens, 0.88 for sick chickens, and 0.92 for healthy chickens, with an
overall mAP of 0.89. Furthermore, the system is integrated with a web dashboard to display detection
results and automated WhatsApp notifications when abnormal conditions are detected. This research
is expected to assist poultry farmers in monitoring chicken health more efficiently and responsively
without the need for continuous direct supervision.

Keywords: YOLOv8, ByteTrack, Object Detection, Poultry Health, OBS, Computer Vision.

Rancang Bangun Sistem Pengering Tembakau Otomatis Berbasis ESP32 Untuk Peningkatan Efisiensi Produksi

2026-02-06T13:51:00+00:00

obacco drying is a critical process that determines the final quality of tobacco leaves. Conventional drying methods still widely used have several drawbacks, including dependence on weather conditions, long drying time, and inconsistent quality. This study aims to develop a smart tobacco drying system based on an ESP32 microcontroller integrated with Internet of Things (IoT) technology to automatically control temperature, humidity, and dryness level. The system uses a DHT22 sensor to measure temperature and humidity, a TCS230 color sensor to detect changes in leaf color, and relays to control the heater and blower. Monitoring is performed through the Blynk application. The research method applies a quantitative approach with applied experimentation. The results show that the system maintains temperature at 40–50°C and reduces humidity from 65% to 42% within 3 hours. The system successfully dries 2 kg of tobacco in 3 hours, faster than conventional methods.

Kata kunci: Sistem Pengeringan Cerdas, Tembakau, Internet of Things (IoT), Optimalisasi Produksi.