<![CDATA[Abstract— This research paper reviews the use of Convolutional Neural Networks (CNNs) to categorize diverse sand type using microscopic images, with an objective of improving quality control in construction materials. The paper compares three CNN architectures—LeNet-5, Inception v3, and ResNet50—for discriminating between specific sand categories, such as two river sands (Cipongkor and Citarum) and three types of silica sand (brown, cream, and white). Each model was trained and tested on different dataset splits, with images pre-processed to highlight specific microscopic properties. To achieve a thorough comparison, each model's performance was measured using a variety of measures such as F1-score, accuracy, recall, and precision. These measurements enabled a comprehensive evaluation of how accurately and reliably each CNN model categorized the various sand types. ResNet50 consistently delivered the highest accuracy, achieving perfect classification in some instances, showcasing its effectiveness in capturing fine details in sand textures. These results highlight the potential of CNN-based approaches for precise and automated sand classification, which supports increased quality assurance in construction and related areas. Index Terms— Convolutional Neural Network (CNN); sand classification; LeNet-5; Inception v3; ResNet50]]>
Copyrights © 2024
