Jurnal Sistem Cerdas

Real-Time Face Age Detection System Based on Deep Neural Networks with MediaPipe Optimization for Enhanced Accuracy

muhaimin iskandar, Nur Azizah, Firman Jaya — Wed, 24 Dec 2025 14:38:28 +0700

The transformation of machine learning and computer vision technology enables computers to automatically learn complex visual patterns, forming the foundation for biometric applications such as identity authentication, face detection, and demographic analytics. Face age estimation predicts age based on facial characteristics in digital images with high accuracy. Handcrafted feature-based approaches such as Histogram of Oriented Gradients (HOG) and Local Binary Patterns (LBP) are less stable against variations in lighting, camera orientation, and facial expressions. Deep learning, particularly Deep Neural Networks (DNN), improves accuracy through automatic hierarchical feature extraction. However, raw image-based methods have high computational loads and require large GPUs, which are less than ideal for real-time use on limited devices. This research proposes a DNN-based age estimation system optimized through MediaPipe Face Mesh geometric features. The system consists of five layers: input, feature extraction (468 facial landmarks), optimization with Principal Component Analysis (PCA) for 64 features, DNN regression (three hidden layers), and output. A custom dataset of 1,235 facial images (ages 3–40 years) was divided into 80% training and 20% testing. The model was trained with the Adam optimizer (learning rate 0.001, epochs 500, loss MAE). Evaluation results: MAE 0.56 years, RMSE 1.94 years, R² 0.9726. Tolerance accuracy: 91% (±1 year), 96.7% (±2 years), 97.5% (±3 years), 99.2% (±5 years). An efficient system for real-time use on low-computing devices, supporting biometric applications such as security, content filtering, personalization, and health. This research contributes to accurate, lightweight, and adaptive age estimation systems.

The Effect of Lighting Variations on the Accuracy of Formalin Detection in Milkfish Using HSV Color Space and k-Nearest Neighbors (kNN) Algorithm

Wed, 24 Dec 2025 15:26:10 +0700

Milkfish (Chanos chanos) is a widely consumed fish commodity in Indonesia, often subject to preservation using formalin, a chemical with serious health risks when misused. This study proposes a non-destructive formalin detection method using HSV (Hue, Saturation, Value) color features extracted from eye images of milkfish, classified via the k-Nearest Neighbor (kNN) algorithm. The research investigates the impact of varying illumination levels low, medium, and high on the consistency of HSV features and the accuracy of kNN classification. Results show that medium lighting conditions yield the highest classification accuracy, suggesting an optimal illumination range for field deployment. The system's simplicity and potential for real-time implementation on mobile or embedded platforms make it suitable for use by non-technical personnel in traditional markets. Challenges such as environmental temperature, image angle, and surface reflectivity are addressed through calibration strategies and operational guidelines. This study contributes practical insights into lighting control and feature stability, enhancing the reliability of image-based formalin detection systems.

Elite-Refined Genetic Algorithm with Hill Climbing Local Search for University Course Scheduling

Wed, 24 Dec 2025 15:13:59 +0700

Abstract— This paper proposes a hybrid optimization approach combining Genetic Algorithm (GA) and Hill Climbing (HC) to address the university course scheduling problem in the Informatics Study Program at Universitas Islam Negeri Siber Syekh Nurjati Cirebon. The hybrid GA-HC model integrates GA’s global exploration capability with HC's local refinement strategy to minimize hard and soft constraint violations while achieving balanced timetables. The dataset includes 56 course classes, 18 lecturers, and three rooms, with scheduling over five working days and 11 time slots per day. Experimental results demonstrate that GA-HC outperforms pure GA and pure HC in convergence speed, average fitness, and stability of feasible solutions. Parameter tuning analysis further shows that moderate mutation rates and limited HC iterations yield optimal trade-offs between runtime and solution quality. The proposed hybrid framework effectively enhances convergence, reduces conflicts, and improves overall timetable quality, confirming its robustness for large-scale academic scheduling problems.

Comparison of SVM and Naive Bayes in Public Sentiment Analysis on Budget Efficiency

Yusmita, Farid Wajidi, Muh.Rafli Rasyid — Wed, 24 Dec 2025 15:23:45 +0700

Kebijakan efisiensi anggaran melalui Instruksi Presiden Nomor 1 Tahun 2025 memicu beragam respons publik di media sosial, khususnya X. Penelitian ini mengklasifikasikan sentimen publik menggunakan algoritma Naïve Bayes dan SVM dengan 6.596 twit setelah tahap praproses, menggunakan pelabelan Lexicon InSet, dan ekstraksi fitur TF-IDF. Hasilnya menunjukkan bahwa SVM-LinearSVC mencapai akurasi tertinggi sebesar 94%, sementara Naïve Bayes mencapai 86% tetapi lebih cepat dalam pelatihan dan prediksi. Temuan ini menegaskan bahwa algoritma pembelajaran mesin efektif untuk memetakan opini publik terkait kebijakan, sekaligus menjadi referensi penelitian analisis sentimen berbahasa Indonesia.

Hand Gesture Detection Implemented based on Long Short-Term Memory (LSTM) Method

Sat, 06 Dec 2025 00:00:00 +0700

The Indonesian government encourages accessibility of information that is friendly to people with disabilities, one of which is through the development of information and communication technology. Efforts to increase accessibility of information and encourage independence of people with disabilities need to be supported by the right solutions. According to the Central Statistics Agency, there were 0.68% of the total population of Indonesia in 2019, this data shows that deafness is one of the highest disabilities in Indonesia. Efforts to increase accessibility of information and encourage independence of people with disabilities need to be supported by the right solutions. One potential solution is the development of a self-service system that is friendly to the deaf. This study aims to develop a self-service system that is friendly to the deaf and helps in obtaining information and services independently. The results achieved in this study are in the application of hand signal detection using the Long Short-Term Memory method which can overcome the problem of long-distance dependency and improve performance in recognizing complex hand signal patterns. The hand signal recognition feature can be improved by overcoming the problem of long-distance dependency with a maximum user distance of 1.25 meters, the system can still recognize hand signals well. It is hoped that in the future, more in-depth studies can be carried out on long-distance dependency for variations of other hand signal recognition methods, so that people with disabilities can more easily use the self-service system.

The Role of Artificial Intelligence in Building a Culture of Knowledge Sharing among Students and University Students

Suarni Norawati, Akmal Andri Yantama, M. Zacky — Mon, 29 Dec 2025 09:49:45 +0700

This study examines the role of Artificial Intelligence (AI) in fostering a culture of knowledge sharing and enhancing knowledge management among students and university learners in the digital learning environment. Using a quantitative explanatory approach, data were collected from 100 respondents who actively used AI-based applications such as ChatGPT, Grammarly, and Copilot in their academic activities. The data were analyzed using Partial Least Squares–Structural Equation Modeling (PLS-SEM) with SmartPLS 4.0. The results reveal that AI has a significant direct effect on knowledge management (β = 0.316, p = 0.000) and a strong positive influence on knowledge sharing behavior (β = 0.851, p = 0.000). Furthermore, knowledge sharing significantly mediates the relationship between AI and knowledge management (β = 0.611, p = 0.000), indicating that AI’s greatest impact occurs through the enhancement of collaborative knowledge exchange among learners. The model explains 80.3% of the variance in knowledge management and 72.3% in knowledge sharing, demonstrating strong predictive power. These findings highlight AI’s potential as a collaborative catalyst that strengthens human-centered learning ecosystems. The study contributes both theoretically by extending the understanding of AI-mediated knowledge processes—and practically by providing insights for educators to integrate AI ethically and effectively into knowledge-based learning systems

A Multivariate LSTM Approach for Monthly Rice Production Forecasting in East Java

Tue, 30 Dec 2025 00:00:00 +0700

Accurate forecasting of rice output is essential for improving regional food security planning, particularly in East Java Province, which serves as a major national rice granary. This study develops a Long Short-Term Memory (LSTM) model to predict rice production using monthly data on production and harvested area from 2018 to 2024. The methodology includes data preprocessing, normalization, sequence construction with a sliding window, training of a multivariate LSTM model, and performance evaluation using mean absolute error (MAE), root mean square error (RMSE), and mean absolute percentage error (MAPE). Results show that the LSTM model achieves superior predictive accuracy, with an MAE of 95,030.16, RMSE of 120,229.01, and MAPE of 16.64%, significantly outperforming baseline Moving Average and Linear Regression models. While the model effectively captures seasonal production trends, some inaccuracies remain during periods of anomalous production values. These findings suggest that the LSTM model is effective for projecting rice production and may provide a foundation for early warning systems and regional food distribution strategies. Further improvements could be realized by integrating climate variables or adopting a hybrid model architecture to enhance predictive precision.

Melanin-Aware and ArcFace Methods in Facial Recognition for Dark-Skinned Individuals

Wed, 31 Dec 2025 11:14:50 +0700

The facial recognition system employing the Melanin-Aware method in conjunction with ArcFace, trained on a dataset of 1,000 dark-skinned facial samples, demonstrates the ability to accurately recognize individuals with dark skin while maintaining performance for non-dark-skinned individuals. ArcFace is utilized as the primary feature extractor, leveraging the additive angular margin to enhance inter-class separability. The experiments were conducted using 1,000 dark-skinned facial samples for training and 100 samples for testing. Evaluation results indicate that melanin-aware preprocessing improves average accuracy by up to 17% compared to the absence of preprocessing, and by 7% compared to the standard aggressive CLAHE-based method. Furthermore, the True Acceptance Rate (TAR) increased from 88.15% to 93.33% at FAR = 1e−2, and from 83.7% to 86.67% at FAR = 1e−3, signifying enhanced system stability under stringent security conditions. The performance gains are supported by a more stable distribution of similarity scores and lower threshold values, reflecting improved separation between genuine and impostor pairs.

Automated Waste Classification for Sustainable Cities Using YOLO Based CNN Integrated IoT

Waluyo Nugroho, Adnan Alfattah, Mada Jimmy Fonda Arifianto, Aswan Hadi — Sun, 04 Jan 2026 00:00:00 +0700

Sustainable waste management is a vital component of smart city development, directly impacting environmental quality and recycling efficiency. This study presents an IoT-enabled waste classification system that utilizes a Convolutional Neural Network (CNN) for accurate, real-time identification of organic and non-organic waste. The model, implemented using the YOLO architecture, was trained on a diverse dataset of waste images captured under various environmental conditions to ensure robustness in practical scenarios. Classification results are automatically stored in a MySQL database and visualized via an Internet of Things (IoT) based Node-RED dashboard, enabling municipal operators to monitor waste categories and quantities remotely. Field evaluations demonstrate that the system achieves an accuracy of 94%, precision of 94.5%, recall of 93.2%, and an F1-score of 93.85%, indicating high detection reliability and consistent performance, even in challenging urban environments. By integrating CNN-based deep learning with IoT visualization tools, this approach offers a scalable and efficient solution that supports sustainable waste management initiatives within smart city frameworks.

WebSocket-Based Smart Surveillance Camera for Real-Time Detection of Occupational Health and Safety PPE Non-Compliance in Industrial Areas

Mon, 05 Jan 2026 00:00:00 +0700

In industrial settings, ensuring adherence to Occupational Health and Safety (OHS) Personal Protective Equipment (PPE) regulations continues to be a crucial challenge. The creation of a WebSocket-based smart surveillance camera system for the real-time identification and reduction of PPE infractions is discussed in the paper. The proposed system includes an ESP32-S3 microcontroller accompanied by an OV5640 camera module, acting as an edge-processing embedded platform. The Edge Impulse machine learning framework was used to train image classification and detection models, enabling efficient low-latency inference directly on the device. A websocket enabled web server streams video frames in real time for constant monitoring, with instant display using regular browsers without wasting bandwidth. Experimental results demonstrate that even with limited computational resources, the system is able to perform on-device inference with very high responsiveness and good detection accuracy. This technology provides a scalable and affordable way to enhance OHS compliance monitoring in industry, reduce reliance on manual supervision, and encourage proactive risk mitigation methodologies.