COMPARATIVE ANALYSIS OF MULTI-CLASS SEMANTIC SEGMENTATION MODELS FOR PIPELINE CORROSION DETECTION

Samosir, David Mario Yohanes (2025) COMPARATIVE ANALYSIS OF MULTI-CLASS SEMANTIC SEGMENTATION MODELS FOR PIPELINE CORROSION DETECTION. Undergraduate thesis, Universitas Pendidikan Ganesha.

[img] Text (COVER)
2115101055-COVER.pdf
Download (1MB)
[img] Text (ABSTRAK)
2115101055-ABSTRAK.pdf
Download (133kB)
[img] Text (BAB 1 PENDAHULUAN)
2115101055-BAB 1 PENDAHULUAN.pdf
Download (212kB)
[img] Text (BAB 2 KAJIAN TEORI)
2115101055-BAB 2 KAJIAN TEORI.pdf
Restricted to Repository staff only
Download (404kB) | Request a copy
[img] Text (BAB 3 METODELOGI PENELITIAN)
2115101055-BAB 3 METODELOGI PENELITIAN.pdf
Restricted to Repository staff only
Download (346kB) | Request a copy
[img] Text (BAB 4 HASIL DAN PEMBAHASAN)
2115101055-BAB 4 HASIL DAN PEMBAHASAN.pdf
Restricted to Repository staff only
Download (2MB) | Request a copy
[img] Text (BAB 5 PENUTUP)
2115101055-BAB 5 PENUTUP.pdf
Restricted to Repository staff only
Download (129kB) | Request a copy
[img] Text (DAFTAR PUSTAKA)
2115101055-DAFTAR PUSTAKA.pdf
Download (130kB)
[img] Text (LAMPIRAN)
2115101055-LAMPIRAN.pdf
Download (509kB)

Abstract

Corrosion detection in pipeline infrastructure necessitates a precise and effective methodology to facilitate preventative maintenance and reduce the likelihood of system failure. Corrosion is material damage due to environmental reactions that require accurate visual detection. Multi-class semantic segmentation enables precise classification of pixels into categories such as corrosion, pipe, and background. Mobile U-Net was selected for its integration of the robust U-Net architecture in intricate object segmentation with the efficiency of MobileNet, whilst BiSeNetV3 is designed for inference speed via its bilateral structure and EfficientNetB1 backbone, enhancing generalization on constrained datasets. The dataset consists of 112 original images enriched to 2,028 images through traditional data augmentation techniques, with a data distribution of 80% for training, 10% for validation, and 10% for testing. In the implementation of the original dataset, Mobile U-Net with a model size of 28.3 MB showed a validation accuracy of 80.58% and reached a mIoU value of 62.03% in the 36th epoch, with a mIoU distribution of 58.67% for the corrosion class, 50.64% for the pipe class, and 76.77% for the background class. BiSeNetV3, with a model size of 13.56 MB, achieved a validation accuracy of 83.96% with a mIoU of 62.01% in the 40th epoch. The application of data augmentation significantly improves both models' performance, with BiSeNetV3 achieving a validation accuracy of 97.54% and mIoU of 90.95% with a computing time of 0.4788 seconds per image, while Mobile U-Net reaches 96.76% accuracy and 88.20% mIoU in 0.4287 seconds. Testing results demonstrate BiSeNetV3's superior performance with a mIoU of 96.12% and a Dice score of 97.91%, compared to Mobile U-Net's 92.41% mIoU and 95.72% Dice score. The system implements pixel-based corrosion area calculation on 256x256 pixel images, quantitatively evaluating corrosion damage. Results indicate BiSeNetV3 as the optimal choice for mobile application implementation based on its balanced performance in segmentation accuracy, computational efficiency, and model size while maintaining consistent segmentation across all classes.

Item Type: Thesis (Undergraduate)
Uncontrolled Keywords: Image Segmentation, Multi-Class Segmentation, Deep Learning, Efficient Model, Pipe Corrosion Detection
Subjects: T Technology > T Technology (General)
T Technology > TA Engineering (General). Civil engineering (General)
Divisions: Fakultas Teknik dan Kejuruan > Jurusan Teknik Informatika > Program Studi Ilmu Komputer (S1)
Depositing User: DAVID MARIO YOHANES SAMOSIR
Date Deposited: 25 Apr 2025 03:34
Last Modified: 25 Apr 2025 03:34
URI: http://repo.undiksha.ac.id/id/eprint/23965

Actions (login required)

View Item View Item