This research is a deep learning-driven assistive device specifically engineered to support visually impaired individuals in identifying Philippine New Generation Series (NGS) peso coins. The primary goal of this study is to facilitate independent and accurate coin recognition through an accessible and user-friendly auditory feedback system. The system leverages computer vision and artificial intelligence to distinguish between various denominations based on their visual features. To train deep learning models, a custom dataset consisting of 1,419-coin images was manually collected, ensuring diversity in coin orientation, lighting conditions, and background environments to reflect real-world usage. Three deep learning architectures-Convolutional Neural Network (CNN), ResNet50, and VGG16-were selected for evaluation, given their proven efficacy in image classification tasks. These models were trained and validated on the prepared dataset, and their performance was compared in terms of accuracy. Among the three, the CNN model outperformed the others, attaining an accuracy of 99.89%, demonstrating its superior capability in recognizing coin types with high precision and minimal error. The finalized CNN model was then integrated into a functional hardware prototype which combines the trained model with a camera module and a microcontroller-based processing unit. Upon capturing an image of the coin, the system classifies its denomination and provides real-time auditory feedback, thereby enabling blind or visually impaired users to accurately identify coins without external assistance.

Public safety has become the main priority nowadays, and one major area of research interest is violence detection. This study shows the development of a deep learning-based monitoring system for real-time violence detection. The audio classification employs Recurrent Neural Network (RNN), particularly Long Short-term Memory (LSTM), having a dataset of 13, 141 audio data gathered using a microphone. The methodology incorporates Data Augmentation, Feature extraction using Mel-frequency Cepstral Coefficients (MFCCS), and Hyperparameter Tuning using Optuna and Hyperband pruning. Evaluation metrics including accuracy, precision score, F1 score, and recall were used to assess the model. The hardware implementation utilizes Jetson Nano, along with a microphone and alarm system for violence detection. The Recurrent Neural Network exhibited the best model, achieving outstanding accuracy of 99.85%. This study depicts the potential of a deep learning-based device in enhancing public security, however, further improvement requires recognizing limitations such as the need for diverse datasets. The study contributes to the increasing potential of AI-powered security systems that offer future advancements in audio-based violence detection.

Semiconductor manufacturing relies on final test (FT) to ensure the quality of integrated circuits before shipment. To optimize manufacturing and reduce costs, this study develops an AI model that can shorten the FT cycle time. Benefits include early detection of problematic lots, more accurate forecasting, and more thorough root cause analysis. Using a Multilayer Perceptron (MLP) to predict lot quality, all programming, training, and evaluation were coded using Python through Jupyter Notebook. Kruskal-Wallis and Dunn's tests showed that increasing hidden layers from two to five did not significantly improve accuracy. Afterward, hyperparameter tuning identified the best model configuration, resulting in to-fold cross-validation with an average classification accuracy of 88.4622%, an F1-score of 88.4309%, and an ROC-AUC of 96.2478%.

Nile tilapia (Oreochromis niloticus) is vital for aquaculture due to its rapid growth and high productivity. This study tackles detection and tracking concerns using machine vision techniques, developing an automated system with DeepLabCut™ and training classifiers with ResNet50 and MobileNet V2-1.0. MobileNet V2 showed balanced performance, with lower training (634.5 pixels) and test errors (723.7 pixels), and a higher mean likelihood (0.60) for the body region. ResNet50, despite efficient learning, suffered from overfitting, with training and test errors of 1121.2 and 1083.3 pixels, and low mean likelihood (0.1) for all body parts. These results indicate MobileNet V2 is better for reliable, resource-efficient predictions, while ResNet50 could improve with better data quality and annotations. Future research should focus on enhancing data quality, extended training, and advanced regularization techniques.

Dissolved copper in water is considered before it can be used for any purpose, especially in the case of Marinduque, Philippines where mine tailing containing high concentrations of the metal was spilled to various water resources. This study aimed to use an optical fiber amplifier to record the color reaction being produced by soluble copper and develop a machine-learning model that determines the level of soluble metals in water. The research utilized a prototype and AI development framework. Preparation of the image processing device and fiber optic sensor (BV501s), the preliminary gathering of baseline data, system modeling & development, and System Evaluation was conducted. The BV501s optical fiber sensor was used as it provides RGB equivalents of tannic reaction to soluble copper. There was a total of 33 samples each with its segmented RGB equivalents. The model for identification of the amount of soluble metal was developed using two (2) models such as multiple regression analysis and support vector regression. Using the applied methodology, copper amounts can be determined in water using regression models. The amount of soluble copper can be identified more accurately using the support vector regression since it yields a higher r2 score compared to the multiple linear regression model using k-fold cross-validation. Integration of the developed model to hardware is also suggested as and comparison of the developed model. Develop a classification model for the potability of the water samples based on the soluble water samples.

The devastation caused by fires is a significant threat to human life. There are traditional fire extinguishing methods but can have negative impacts on the environment. This study utilized data from a system that uses sound waves to extinguish fires without requiring water and chemicals. This paper created machine learning models that can predict whether a fire can be extinguished by the sound waves given the features like the size, fuel, distance, decibel, airflow, and frequency. The researchers used Python programming to create different machine learning models and determined the most accurate model using the classification accuracy and F1 score as performance metrics. The XGBoost model was identified as the most effective in classifying the sound wave flame extinction with accuracy scores of 98.31 % and 98.62% for the model with default and optimized parameters, respectively.

Pineapple's size is very crucial in determining its market value. Size sorting is commonly done via visual inspection, which is usually subject to inconsistency and errors. Errors due to failed sorting may either lead to wastage or loss, or mispricing. This study presents incorporation of the machine learning techniques like Logistic Regression, K-Nearest Neighbors, Support Vector Machine, Decision Tree, and Random Forest in classifying pineapple sizes as small, medium, and large using the extracted features of images processed via OpenCV libraries as well as Python Programming. A total of 300 pineapples of different sizes were captured and processed to extract features such as the area, width, height, enclosed-circle radius, and perimeter. The models were optimized using GridSearchCV and were evaluated using accuracy and F1 score metrics. Based on the results, SVM was found to be the most suited classification model, having an optimized training and testing accuracy of 95.67 % and 96.67 %, respectively, and an F1 score of 96.67 %.

Short Message Service (SMS) is a generally used communication method due to its convenience and affordability. SMS spam message is an unauthorized text message that contains a variety of content types such as advertisements, fraudulent texts, and promotions. These messages can pose a serious threat to mobile phone users as they may contain security threats, malicious activities, and other concerning issues. These can lead to identity theft, financial loss, and other types of fraud. To deal with the problem of spamming, various machine-learning models are applied to develop an optimized model that effectively, reliably, and precisely identifies and filter out spam or junk message from a genuine SMS text. The dataset used is a combination of self-acquired data and internet collected dataset with 60–40 ham to spam partitions. With regards to the accuracy of the model, the Bernoulli Naive Bayes achieved the highest performance with 96.63% accuracy upon optimization.

The lack of a transport quality forecasting system in farming and sorting facilities of indigenous varieties of bananas is aiding the increase of food waste generation in the country. This in turn decreases agriculture sustainability and imposes economic losses to farmers. Musa acuminata ‘Señorita’ are diploid cultivars of bananas originating in the Philippines. This study aims to develop a machine learning-based system that classifies Musa acuminata ‘Señorita’ bananas into their transport tiers: (I) for interprovincial distribution, (II) for intra provincial distribution, or (III) subject for rejection. The model is trained and went through 7 machine learning classifiers to identify which model is the most compatible with the system design. The application of external parameters such as size, girth, weight, maturity stage, and RGB parameters can be the foundation to develop a machine learning-based banana transport tier classifier that accurately monitors and determines how long they can travel based on their maturity level. Among the seven models, Logistic Regression, Linear Discriminant Analysis, Decision Tree Classifier, Gaussian Naive Bayes, and Support Vector Machine attained a classification accuracy of 100 %. The development of this system can aid in the proper distribution of produce and help uplift the country's agriculture and economic sustainability.

Water is an essential natural resource for life on Earth, and it is the foundation of all living things. However, water pollution is a growing environmental concern caused by human activities, such as improper waste disposal and the discharge of untreated sewage. The consequences of this problem on human health and aquatic life highlight the need for effective supervision and administration of water reserves. This research paper aims to utilize a machine learning approach to predict water quality and identify the most influential features affecting water potability. These features were obtained from three methods, namely Univariate Selection, Recursive Feature Elimination, and Feature Importance, to identify the most influential features. The study compares the performance of various classification algorithms, including K-Nearest Neighbor, Decision Tree, Random Forest, AdaBoost, XGBoost, Linear Discriminant Analysis, Gaussian Naïve Bayes, Logistic Regression, MLPClassifier, and ExtraTree Classifier, using evaluation criteria such as accuracy, precision, recall, F1 score, and computational efficiency. After conducting all these processes, ExtraTree Classifier achieved the highest accuracy of 89 % among the compared machine learning models. Overall, the results of this research may contribute to better public health outcomes and improved management of water resources.

Smart farming system using necessary infrastructure is an innovative technology that helps improve the quality and quantity of agricultural production in the country including tomato. Since tomato plant farming take considerations from various variables such as environment, soil, and amount of sunlight, existence of diseases cannot be avoided. The recent advances in computer vision made possible by deep learning has paved the way for camera-assisted disease diagnosis for tomato. This study developed the innovative solution that provides efficient disease detection in tomato plants. A motor-controlled image capturing box was made to capture four sides of every tomato plant to detect and recognize leaf diseases. A specific breed of tomato which is Diamante Max was used as the test subject. The system was designed to identify the diseases namely Phoma Rot, Leaf Miner, and Target Spot. Using dataset of 4,923 images of diseased and healthy tomato plant leaves collected under controlled conditions, we train a deep convolutional neural network to identify three diseases or absence thereof. The system used Convolutional Neural Network to identify which of the tomato diseases is present on the monitored tomato plants. The F-RCNN trained anomaly detection model produced a confidence score of 80 % while the Transfer Learning disease recognition model achieves an accuracy of 95.75 %. The automated image capturing system was implemented in actual and registered a 91.67 % accuracy in the recognition of the tomato plant leaf diseases.