Concrete Surface Crack Detection via YOLOv8 Deep Learning Algorithm

Ahmet Hakan Altun

Makale detayları

Concrete Surface Crack Detection via YOLOv8 Deep Learning Algorithm

Yazarlar:

Ahmet Hakan Altun

Yayıncı:

Cukurova University

Ana Başlıklar:

Yapı Malzemeleri

Dergi:

Çukurova Üniversitesi Mühendislik Fakültesi Dergisi Vol.39 Issue 3, pp.667-678

Belge Tipi:

Makale

Yayın Tarihi:

3/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001329024

pISSN:

2757-9255

Anahtar kelimeler:

Özet:

 Concrete should be monitored throughout its service life, any damages should be detected, and necessary repairs should be promptly carried out. Therefore timely and accurate detection is crucial for the durability of concrete. Cracks are the earliest indicators of damage in reinforced concrete structures. Especially in high seismic risk regions like Turkey, early detection of cracks is of vital importance for the resilience and safety of structures. Manual detection of cracks is generally disadvantaged in terms of time, labor, cost, high error probability, and application difficulties. As an alternative to manual inspection, image processing techniques and algorithms based on machine learning and deep learning are increasingly being utilized in this field. This study aims to detect cracks on concrete surfaces using image processing methods with the METU dataset consisting of images from various buildings on the Middle East Technical University campus. A total of 550 sample images were selected from the dataset, comprising 500 positive and 50 negative images. The dataset was expanded to 1330 examples using various data augmentation techniques. The dataset was divided into 88 training, 8 validation, and 4 test sets. Thus 1170 images were used for training, 105 for validation, and 55 for testing. The training process was conducted in the Google Colab environment using the YOLOv8 model from the YOLO series. According to the results obtained, the model produced very few false positive results in crack predictions and demonstrated high accuracy in distinguishing different classes.

 Concrete should be monitored throughout its service life, any damages should be detected, and necessary repairs should be promptly carried out. Therefore timely and accurate detection is crucial for the durability of concrete. Cracks are the earliest indicators of damage in reinforced concrete structures. Especially in high seismic risk regions like Turkey, early detection of cracks is of vital importance for the resilience and safety of structures. Manual detection of cracks is generally disadvantaged in terms of time, labor, cost, high error probability, and application difficulties. As an alternative to manual inspection, image processing techniques and algorithms based on machine learning and deep learning are increasingly being utilized in this field. This study aims to detect cracks on concrete surfaces using image processing methods with the METU dataset consisting of images from various buildings on the Middle East Technical University campus. A total of 550 sample images were selected from the dataset, comprising 500 positive and 50 negative images. The dataset was expanded to 1330 examples using various data augmentation techniques. The dataset was divided into 88 training, 8 validation, and 4 test sets. Thus 1170 images were used for training, 105 for validation, and 55 for testing. The training process was conducted in the Google Colab environment using the YOLOv8 model from the YOLO series. According to the results obtained, the model produced very few false positive results in crack predictions and demonstrated high accuracy in distinguishing different classes.

 Concrete should be monitored throughout its service life, any damages should be detected, and necessary repairs should be promptly carried out. Therefore timely and accurate detection is crucial for the durability of concrete. Cracks are the earliest indicators of damage in reinforced concrete structures. Especially in high seismic risk regions like Turkey, early detection of cracks is of vital importance for the resilience and safety of structures. Manual detection of cracks is generally disadvantaged in terms of time, labor, cost, high error probability, and application difficulties. As an alternative to manual inspection, image processing techniques and algorithms based on machine learning and deep learning are increasingly being utilized in this field. This study aims to detect cracks on concrete surfaces using image processing methods with the METU dataset consisting of images from various buildings on the Middle East Technical University campus. A total of 550 sample images were selected from the dataset, comprising 500 positive and 50 negative images. The dataset was expanded to 1330 examples using various data augmentation techniques. The dataset was divided into 88 training, 8 validation, and 4 test sets. Thus 1170 images were used for training, 105 for validation, and 55 for testing. The training process was conducted in the Google Colab environment using the YOLOv8 model from the YOLO series. According to the results obtained, the model produced very few false positive results in crack predictions and demonstrated high accuracy in distinguishing different classes.

 Beton kullanım ömrü boyunca takip edilmeli, varsa hasarlar tespit edilmeli ve gerekli işlemler zamanında yapılmalıdır. Bundan dolayı doğru zamanda doğru tespit betonun dayanıklılığı açısından oldukça önemlidir. Çatlaklar, betonarme yapıların zarar gördüğünün en erken sinyalleridir. Türkiye gibi deprem riski yüksek bölgelerde yapıların dayanıklılığı ve güvenliği açısından çatlakların erken tespiti hayati öneme sahiptir. Çatlakları manuel olarak tespit etmek genellikle zaman, işgücü, maliyet, yüksek hata olasılığı ve uygulamadaki zorluklar açısından oldukça dezavantajlıdır. Manuel tespite alternatif olarak görüntü işleme teknikleri, makine öğrenmesi ve derin öğrenme tabanlı algoritmaların bu alanda kullanımı yaygınlaşmaktadır. Bu çalışmada, Orta Doğu Teknik Üniversitesi kampüsündeki farklı binalardan elde edilen görüntülerden oluşan METU veri kümesi kullanılarak beton yüzeyindeki çatlakların görüntü işleme yöntemi ile tespit edilmesi amaçlanmıştır. Veri kümesinden 550 adet örnek görüntü seçilmiş olup bu görüntülerin 500 adedi pozitif, kalan 50 adedi ise negatif görüntüden oluşmaktadır. Veri seti çeşitli veri artırma teknikleri ile 1330 örneğe genişletilmiştir. Veri seti 88 eğitim, 8 doğrulama, 4 test kümesi olarak bölünmüştür. Sonuç olarak 1170 adet görüntü eğitim, 105 adet görüntü doğrulama ve 55 adet görüntü ise test için kullanılmıştır. Eğitim işlemi Google Colab ortamında gerçekleştirilmiştir. Model olarak YOLO serisinden YOLOv8 modeli kullanılmıştır. Elde edilen sonuçlara göre modelin çatlak tahminlerinde çok az yanlış pozitif sonuç verdiği ve farklı sınıfları ayırt etmede yüksek başarı gösterdiği tespit edilmiştir.

Makale detayları

Concrete Surface Crack Detection via YOLOv8 Deep Learning Algorithm

Yazarlar:

Ahmet Hakan Altun

Yayıncı:

Cukurova University

Ana Başlıklar:

Yapı Malzemeleri

Dergi:

Çukurova Üniversitesi Mühendislik Fakültesi Dergisi Vol.39 Issue 3, pp.667-678

Belge Tipi:

Makale

Yayın Tarihi:

3/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001329024

pISSN:

2757-9255

Concrete

Crack detection

Image processingBeton

Çatlak tespiti

Görüntü işleme

Sayı

Makale detayları

Concrete Surface Crack Detection via YOLOv8 Deep Learning Algorithm

Yazarlar: Ahmet Hakan Altun

Yayıncı: Cukurova University

Ana Başlıklar: Yapı Malzemeleri

Dergi: Çukurova Üniversitesi Mühendislik Fakültesi Dergisi Vol.39 Issue 3, pp.667-678

Belge Tipi: Makale

Yayın Tarihi: 3/2024

Belge Dili: Türkçe

Document ID (Org No.): 2001329024

pISSN: 2757-9255

Concrete

Crack detection

Image processingBeton

Çatlak tespiti

Görüntü işleme

Sayı

Yazarlar:

Ahmet Hakan Altun

Yayıncı:

Cukurova University

Ana Başlıklar:

Yapı Malzemeleri

Dergi:

Çukurova Üniversitesi Mühendislik Fakültesi Dergisi Vol.39 Issue 3, pp.667-678

Belge Tipi:

Makale

Yayın Tarihi:

3/2024

Belge Dili:

Türkçe

Document ID (Org No.):

2001329024

pISSN:

2757-9255