Claim Missing Document
Check
Articles

Found 3 Documents
Search

Automated Pixel-Level Concrete Defect Detection using U-Net Architecture: A Comparative Study with Clustering-Based Segmentation Hendri, Halifia; Rani, Larissa Navia; Enggari, Sofika; Ramadhanu, Agung; Hadi, Febri
Journal of Applied Data Sciences Vol 7, No 2: May 2026
Publisher : Bright Publisher

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47738/jads.v7i2.1298

Abstract

Concrete surface defect detection is a critical aspect of maintaining the integrity and safety of infrastructure in civil engineering. Traditional manual inspection methods are time-consuming, prone to human subjectivity, and often limited by physical accessibility, necessitating the development of robust automated solutions. This paper presents an automated pixel-level concrete surface defect detection system utilizing the U-Net deep learning architecture. The primary contribution and novelty of our approach lie in optimizing the network's encoder-decoder structure with skip connections to effectively capture both broad contextual features and precise spatial localization. This overcomes the critical limitations of existing traditional methods, which frequently struggle with complex concrete background textures, inherent noise, and uneven illumination. To validate our approach, the proposed U-Net model is systematically compared against a widely used baseline method, K-Means clustering combined with Gray-Level Co-occurrence Matrix (GLCM) texture analysis. The evaluation was conducted using a comprehensive dataset consisting of 1000 high-resolution concrete images. Experimental results reveal that the deep learning architecture vastly outperforms the traditional baseline. Specifically, the U-Net model achieved an outstanding F1-Score of 92.47%, a precision of 93.18%, and a mean Intersection over Union (mIoU) of 86.55%. In stark contrast, the K-Means and GLCM approach only yielded an F1-Score of 69.83% and an mIoU of 54.21%. These quantitative findings demonstrate that the proposed U-Net-based system not only successfully minimizes false segmentations but also provides a highly reliable, efficient, and scalable computational framework. Ultimately, this research delivers a practical solution that can be seamlessly integrated into continuous automated structural health monitoring systems, paving the way for safer and more proactive civil infrastructure management.
A Modified Watershed Algorithm for Rice Plant Growth Stage Analysis Putra, Teri Ade; Yuhandri, Yuhandri; Ramadhanu, Agung
Journal of Applied Data Sciences Vol 7, No 2: May 2026
Publisher : Bright Publisher

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.47738/jads.v7i2.1117

Abstract

Information technology plays a crucial role in enhancing various sectors, including agriculture. In particular, technological advancements in crop monitoring are essential for sustainable food production, where accurate growth analysis is vital. Image-based approaches have emerged as a promising tool for assessing crop growth, particularly in rice plants. This study aims to enhance rice plant image segmentation using an improved Watershed algorithm, integrating the Laplacian operator and Distance Transform. This study utilizes a Support Vector Machine (SVM) classifier for segmenting and classifying rice plant growth stages, achieving accuracy, precision, recall, and F1-score metrics. The dataset consists of 1080 images of rice plants, with 74 images used for training, 31 for testing, and 975 images for validation. The image processing pipeline involves preprocessing steps such as grayscale conversion, normalization, color segmentation, Otsu thresholding, filtering, and edge detection. Following preprocessing, the Watershed algorithm is applied in two scenarios: the conventional method and the enhanced method with the Laplacian operator and Distance Transform. Performance evaluation is based on accuracy, precision, recall, and F1-score metrics. The results show that the enhanced Watershed algorithm significantly outperforms the conventional method, achieving an accuracy of 99.58%, precision of 80.55%, recall of 79.92%, and an F1-score of 81.50%. While there is a slight imbalance in precision and recall, the model demonstrates reliable performance in identifying rice plant growth. This study confirms that integrating the Laplacian operator and Distance Transform into the Watershed algorithm significantly improves segmentation accuracy, supporting the development of automated monitoring systems in smart farming. Furthermore, this approach opens avenues for application in other crops and diverse environmental conditions.
Klasifikasi Aksesori Fashion Berdasarkan Fitur Citra Menggunakan K-Means Clustering Tomi, Zebbil Billian; Ramadhanu, Agung
Intellect : Indonesian Journal of Learning and Technological Innovation Vol. 4 No. 02 (2025): Intellect : Indonesian Journal of Learning and Technological Innovation
Publisher : Yayasan Lembaga Studi Makwa

Show Abstract | Download Original | Original Source | Check in Google Scholar | DOI: 10.57255/intellect.v4i02.1476

Abstract

The rapid development of computer vision and machine learning has enabled new applications in the fashion industry, particularly in image-based product classification and recommendation systems. This study aims to classify fashion accessories, namely wallets, bags, and belts, based on image features using the K-Means clustering algorithm. The dataset consists of 30 images acquired under controlled conditions with uniform lighting, resolution, and background. Although the dataset size is relatively limited, this study is designed as an initial baseline to evaluate the effectiveness of K-Means clustering on small and homogeneous datasets, which are commonly encountered in early-stage image classification research. The research workflow includes image preprocessing (resizing, color space conversion, and noise reduction), object segmentation, and feature extraction focusing on color, texture, and shape characteristics. The extracted features include Local Binary Pattern (LBP), entropy, edge density, eccentricity, extent, and area ratio. The results demonstrate that K-Means clustering is capable of grouping fashion accessories into distinct categories according to their visual characteristics. From a practical perspective, the proposed approach can be applied to automated fashion product cataloging to support inventory management, image-based product search, and recommendation systems in e-commerce platforms. This study provides a simple and interpretable baseline for fashion accessory classification and serves as a foundation for future work involving larger datasets, advanced feature descriptors, or deep learning-based methods. Abstrak Perkembangan computer vision dan machine learning memungkinkan penerapan baru dalam industri fesyen, khususnya pada sistem klasifikasi dan rekomendasi produk berbasis citra. Penelitian ini bertujuan mengklasifikasikan aksesori fesyen berupa dompet, tas, dan ikat pinggang berdasarkan fitur citra menggunakan algoritme K-Means clustering. Dataset yang digunakan terdiri dari 30 citra yang dikumpulkan dalam kondisi terkontrol dengan pencahayaan, resolusi, dan latar belakang seragam. Meskipun jumlah dataset relatif terbatas, pendekatan ini dirancang sebagai studi awal (baseline) untuk mengevaluasi efektivitas K-Means pada dataset kecil dan homogen yang umum dijumpai pada tahap awal pengembangan sistem klasifikasi berbasis citra. Tahapan penelitian meliputi preprocessing (penyeragaman ukuran, konversi warna, dan reduksi noise), segmentasi objek, serta ekstraksi fitur warna, tekstur, dan bentuk. Fitur yang digunakan meliputi Local Binary Pattern (LBP), entropi, kerapatan tepi, eksentrisitas, extent, dan rasio area. Hasil penelitian menunjukkan bahwa algoritme K-Means mampu mengelompokkan aksesori fesyen ke dalam kategori yang berbeda berdasarkan karakteristik visualnya. Secara praktis, hasil penelitian ini berpotensi diterapkan sebagai sistem klasifikasi otomatis pada katalog produk fesyen digital untuk mendukung manajemen inventori, pencarian produk berbasis citra, serta sistem rekomendasi pada platform e-commerce. Penelitian ini diharapkan dapat menjadi baseline sederhana dan interpretatif dalam klasifikasi aksesori fesyen, serta menjadi pijakan untuk pengembangan lanjutan menggunakan dataset yang lebih besar, deskriptor fitur modern, maupun metode berbasis deep learning.