YOLOv8 MODEL OPTIMIZATION WITH DUAL ATTENTION FOR CONCRETE CRACK SEGMENTATION

Kautsar, Isna Rafif (2026) YOLOv8 MODEL OPTIMIZATION WITH DUAL ATTENTION FOR CONCRETE CRACK SEGMENTATION. Tugas Akhir thesis, University of Technology Yogyakarta.

Text 5220411196_Isna Rafif Kautsar_ABSTRAK.pdf Download (145kB)

Abstract

ABSTRACT The development of information technology has significantly impacted the maintenance of building and road infrastructure, particularly in the inspection of concrete cracks. Traditionally, concrete crack identification has been performed manually; however, this approach has several limitations in terms of time, cost, and accuracy. To address these challenges, computer vision-based deep learning techniques, especially Convolutional Neural Networks (CNNs), are increasingly being employed. Among the latest models recognized for their superior efficiency and accuracy in concrete crack segmentation tasks is YOLOv8. This study aims to optimize the YOLOv8 architecture specifically for concrete crack segmentation. In the segmentation results of the standard YOLOv8-seg model, the mAP@50-95 metric is 0.242, representing a 7.5% improvement over YOLOv7-seg. To increase the model’s sensitivity to capturing concrete crack details, this study also integrates several attention mechanisms, including Efficient Channel Attention (ECA), Shuffle Attention (SA), Global Attention Mechanism (GAM), and ResBlock CBAM (Convolutional Block Attention Module). The results show that, for concrete crack segmentation, the attention mechanism in YOLOv8 produces significant performance improvements over standard models with GAM and ResBlock-CBAM. This mechanism enhances the model’s ability to recognize finer crack details while preserving the image’s global context. Tests show that the YOLOv8-seg architecture with Dual Attention increases mAP@50 to 0.757 (a 5.7% improvement) and mAP@50-95 to 0.258 (a 6.6% improvement) compared to the standard YOLOv8 model, achieving the highest results among all models tested in this study. This approach enables fast and accurate mapping of concrete cracks, enabling more effective decision-making in building infrastructure maintenance. Keywords: deep learning, computer vision, YOLOv8 instance segmentation, attention mechanism, concrete crack segmentation

Item Type:	Thesis (Skripsi, Tugas Akhir or Kerja Praktek) (Tugas Akhir)
Subjects:	T Technology > T Technology (General) > T201 Patents. Trademarks
Divisions:	Fakultas Sains Dan Teknologi > S1 Informatika
Depositing User:	Kaprodi S1 Informatika UTY
Date Deposited:	06 May 2026 08:06
Last Modified:	06 May 2026 08:06
URI:	http://eprints.uty.ac.id/id/eprint/19805

Actions (login required)

View Item