Comparative analysis of encoder variants in deep learning-based semantic segmentation of concrete cracks

Hasan Polat; Serhat Alpergin; Mehmet Siraç Özerdem

Article Details

Comparative analysis of encoder variants in deep learning-based semantic segmentation of concrete cracks

Authors:

Hasan Polat | Serhat Alpergin | Mehmet Siraç Özerdem

Publisher:

Dicle University

Main Topics:

Derin Öğrenme

Journal:

Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi Vol.15 Issue 3, pp.581-593

Document Type:

Article

Publication Date:

9/2024

Document Language:

Turkish

Document ID (Org No.):

2001328165

pISSN:

1309-8640

Keywords:

Abstract:

 Following natural disasters such as earthquakes, floods, and fires, significant damages manifest in both buildings and urban infrastructure. Cracks are widely recognized as the predominant indicators of damage or deterioration in concrete structures. Hence, the early and accurate detection of crack defects is crucial to ensure structural safety and longer service life. Deep learning architectures, which have made a significant breakthrough in computer vision applications in recent years, have begun to be widely used in the automatic detection and segmentation of concrete cracks. In particular, deep learning-based segmentation architectures, typically comprising an encoder and a decoder part, play a crucial role in conducting thorough structural health analyses by precisely detecting cracks along with their spatial boundaries. However, encoder block limitations such as the small receptive field of convolution kernels, information losses caused by the pooling operation, and insufficient local feature processing can hinder segmentation performance. This study examines the efficacy of various backbone architectures ResNet-18, ResNet-50, MobileNetV2, Xception, and Inception-ResNet as employed in the encoder block within the DeepLabV3 framework, proposed for the segmentation of cracks on concrete surfaces. The effectiveness of low-level and high-level features provided by different backbone architectures in the encoder part was evaluated on open-access DeepCrack and CrackForest datasets. The results revealed that the MobileNetV2 architecture was the most successful network in terms of learnable parameters and segmentation performance for both data sets. The MobileNetV2 encoder-based segmentation framework achieved 0.81 and 0.70 Dice similarity coefficient DSC for both datasets, respectively, using approximately 6.7 million learnable weights.

 Following natural disasters such as earthquakes, floods, and fires, significant damages manifest in both buildings and urban infrastructure. Cracks are widely recognized as the predominant indicators of damage or deterioration in concrete structures. Hence, the early and accurate detection of crack defects is crucial to ensure structural safety and longer service life. Deep learning architectures, which have made a significant breakthrough in computer vision applications in recent years, have begun to be widely used in the automatic detection and segmentation of concrete cracks. In particular, deep learning-based segmentation architectures, typically comprising an encoder and a decoder part, play a crucial role in conducting thorough structural health analyses by precisely detecting cracks along with their spatial boundaries. However, encoder block limitations such as the small receptive field of convolution kernels, information losses caused by the pooling operation, and insufficient local feature processing can hinder segmentation performance. This study examines the efficacy of various backbone architectures ResNet-18, ResNet-50, MobileNetV2, Xception, and Inception-ResNet as employed in the encoder block within the DeepLabV3 framework, proposed for the segmentation of cracks on concrete surfaces. The effectiveness of low-level and high-level features provided by different backbone architectures in the encoder part was evaluated on open-access DeepCrack and CrackForest datasets. The results revealed that the MobileNetV2 architecture was the most successful network in terms of learnable parameters and segmentation performance for both data sets. The MobileNetV2 encoder-based segmentation framework achieved 0.81 and 0.70 Dice similarity coefficient DSC for both datasets, respectively, using approximately 6.7 million learnable weights.

 Following natural disasters such as earthquakes, floods, and fires, significant damages manifest in both buildings and urban infrastructure. Cracks are widely recognized as the predominant indicators of damage or deterioration in concrete structures. Hence, the early and accurate detection of crack defects is crucial to ensure structural safety and longer service life. Deep learning architectures, which have made a significant breakthrough in computer vision applications in recent years, have begun to be widely used in the automatic detection and segmentation of concrete cracks. In particular, deep learning-based segmentation architectures, typically comprising an encoder and a decoder part, play a crucial role in conducting thorough structural health analyses by precisely detecting cracks along with their spatial boundaries. However, encoder block limitations such as the small receptive field of convolution kernels, information losses caused by the pooling operation, and insufficient local feature processing can hinder segmentation performance. This study examines the efficacy of various backbone architectures ResNet-18, ResNet-50, MobileNetV2, Xception, and Inception-ResNet as employed in the encoder block within the DeepLabV3 framework, proposed for the segmentation of cracks on concrete surfaces. The effectiveness of low-level and high-level features provided by different backbone architectures in the encoder part was evaluated on open-access DeepCrack and CrackForest datasets. The results revealed that the MobileNetV2 architecture was the most successful network in terms of learnable parameters and segmentation performance for both data sets. The MobileNetV2 encoder-based segmentation framework achieved 0.81 and 0.70 Dice similarity coefficient DSC for both datasets, respectively, using approximately 6.7 million learnable weights.

 Depremler, seller ve yangınlar gibi doğal afetler akabinde yapılarda ve kentsel altyapıda ciddi hasarlar meydana gelmektedir. Çatlaklar, beton yapılarda meydana gelen hasarların veya bozulmaların en yaygın belirtileri olarak kabul edilmektedir. Dolayısıyla, çatlak kusurlarının erken ve doğru bir şekilde tespit edilmesi, bu tür yapıların güvenliklerinin sağlanması ve hizmet süreleri açısından önem arz etmektedir. Son yıllarda bilgisayarlı görü uygulamalarında önemli bir atılım sergileyen derin öğrenme mimarileri, beton çatlaklarının otomatik olarak tespit ve segmente edilmesinde yaygın olarak kullanılmaya başlanmıştır. Özellikle, genel olarak bir kodlayıcı ve bir kod çözücü bloktan oluşan derin öğrenme tabanlı segmentasyon mimarileri çatlakları uzamsal sınırları ile tespit ederek, kapsamlı yapı sağlığı analizlerini mümkün kılmaktadır. Ancak, evrişimsel filtrede küçük alıcı alan, pooling işleminin neden olduğu bilgi kayıpları ve yetersiz yerel özellik işlenmesi gibi kodlayıcı blok sınırlandırmaları segmentasyon performansını sekteye uğratmaktadır. Bu çalışmada, beton yüzeylerindeki çatlakların segmentasyonu için önerilen DeepLabV3 mimarisinde kodlayıcı blok için farklı omurga mimarilerinin ResNet-18, ResNet-50, MobileNetV2, Xception ve Inception-ResNet etkinlikleri analiz edilmiştir. Farklı omurga mimariler ile sağlanan alçak ve yüksek seviyeli özelliklerin etkinliklerinin test edilmesi için erişime açık Deepcrack ve CrackForest veri setleri kullanılmıştır. Bulgular her iki veri seti için de MobileNetV2 mimarisinin eğitilebilir parametre ve segmentasyon perfromansı açısından en başarılı ağ olduğunu göstermiştir. MobileNetV2 kodlayıcı tabanlı segmentasyon çerçevesi, yaklaşık 6.7 milyon eğitilebilir ağırlık kullanarak her iki veri seti için sırasıyla 0.81 ve 0.70 Dice benzerlik katsayısı DSC başarımı elde etmiştir.

Article Details

Comparative analysis of encoder variants in deep learning-based semantic segmentation of concrete cracks

Authors:

Hasan Polat | Serhat Alpergin | Mehmet Siraç Özerdem

Publisher:

Dicle University

Main Topics:

Derin Öğrenme

Journal:

Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi Vol.15 Issue 3, pp.581-593

Document Type:

Article

Publication Date:

9/2024

Document Language:

Turkish

Document ID (Org No.):

2001328165

pISSN:

1309-8640

Semantic Segmentation

Deep Learning

Image Processing

Crack Detection

Crack SegmentationSemantik Segmentasyon

Derin Öğrenme

Görüntü İşleme

Çatlak Tespiti

Çatlak Segmenasyonu

Yapı Denetimi

Issues

Article Details

Authors: Hasan Polat | Serhat Alpergin | Mehmet Siraç Özerdem

Publisher: Dicle University

Main Topics: Derin Öğrenme

Journal: Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi Vol.15 Issue 3, pp.581-593

Document Type: Article

Publication Date: 9/2024

Document Language: Turkish

Document ID (Org No.): 2001328165

pISSN: 1309-8640

Issues

Authors:

Hasan Polat | Serhat Alpergin | Mehmet Siraç Özerdem

Publisher:

Dicle University

Main Topics:

Derin Öğrenme

Journal:

Dicle Üniversitesi Mühendislik Fakültesi Mühendislik Dergisi Vol.15 Issue 3, pp.581-593

Document Type:

Article

Publication Date:

9/2024

Document Language:

Turkish

Document ID (Org No.):

2001328165

pISSN:

1309-8640