Crack Segmentation Using DeepLab

Voon, Zhen Cheng (2020) Crack Segmentation Using DeepLab. Final Year Project (Bachelor), Tunku Abdul Rahman University College.

Text Voon Zhen Cheng_Fulltext.pdf Restricted to Registered users only Download (1MB)

Abstract

Crack detection which normally done manually by specialist consume a lot of time and the inspection result might difference depending on the specialist experience and knowledge. Crack that appears in high or narrow place also make crack inspection hard to carry out. In this paper, an automated crack segmentation model which build using DeepLab model with encoder decoder structure is proposed where transfer learning is being utilized in building the model. The model is trained on the dataset from DeepCrack which consists of 300 training images and 237 testing images. 3 models are trained with different value of training step and training rate. The models are then evaluated using the mean intersection-over-union metrics and 3 of the models managed to achieve value around 0.75 for mean intersection-over-union. 10 images also chosen and the precision and recall value for each of the images are calculated and plotted on a graph. The result of the DeepLab model also been used to compare with the result which use Otsu’s method in crack segmentation. This is to compare the difference between crack segmentation which use deep learning and thresholding. Next, limitations of the model are stated which are limited images size, less dataset and unable to segment correctly foreign object. Finally, conclusion is made to summarize the paper.