Abstract
This study presents a system using an image processing technique that
evaluates the pavement condition from an image. Pavement condition
evaluation is an integral part of roads and highway maintenance works,
which mostly depends on human inspection. Although recently some
researches have been conducted on road condition detection with image
processing, these researches used huge databases and deep CNNs that
require expansive computer and longer training time, which limits the
use of deep CNN in practical problems where huge database collection is
not possible always. To solve this problem, in this study, transfer
learning in deep CNN is applied and with only 195 images in each
category, pre-trained VGG-16 and Inception-ResNet v2 models are used for
pavement condition evaluation. VGG-16 achieved more than 90% prediction
accuracy, while Inception-ResNet v2 achieved more than 85% prediction
accuracy. Moreover, to validate the performance, both models have been
tested with random images collected from Google. Evaluating pavement
conditions this way would reduce the need for human inspection. Finally,
the outcome of the study shows that the transfer learning approach could
be useful in research areas, especially in civil engineering, where
image data is insufficient.