EXPLORE PUBLICATIONS BY COUNTRIES

Deep Learning, Convolutional Neural Network, Transfer Learning, Crack Detection.

Computer vision (CV) -based techniques are being deployed to solve the problem of Crack Detection in metallic and concrete surfaces. This is because the Human-oriented inspections being used have drawbacks in the area of cost and manpower. One of the deployed CV techniques is the Deep Convolutional Neural Network (DCNN). Existing DCNN based crack detection models have a challenge of performing poorly when tested on images taken at a different location from the training images, hence crack localization is required. Thus, this research develops a location invariant crack detection and localization (LICDAL) model in unconstrained oil pipeline images using DCNN. LICDAL is developed by applying transfer learning on the Faster Region based - CNN (Faster R-CNN). The model is made location invariant by gathering images of cracked oil pipeline from various locations. The collected images are split into a 70%:30% ratio for training and testing set. LICDAL is evaluated using the mean Average Precision (mAP). The results on testing LICDAL shows the detected and localised cracks with a mAP of 97.3% on a set of 10 new test images taken from different locations; the highest Average Precision at 99% and the lowest Average Precision at 86%. The performance of LICDAL is compared to an existing crack detection model which detects cracks alone. LICDAL adequately localizes the detected cracks, thus improving crack identification. Secondly, there is no drastic reduction in performance for the test images taken at different locations from the training images, thus making LICDAL location invariant.