PROJECT TITLE :

DeepCrack Learning Hierarchical Convolutional Features for Crack Detection

ABSTRACT:

Many computer-vision programmes are attracted to the usual line formations known as cracks. Image-based fracture detection using low-level features is difficult since many cracks (e.g., pavement cracks) have poor continuity and low contrast. DeepCrack is a deep convolutional neural network that can be trained end-to-end to learn high-level characteristics for crack representation in order to automatically detect cracks. The line structures are captured by fusing together multi-scale deep convolutional features obtained at hierarchical convolutional stages. Larger-scale feature maps provide more detailed representations, whereas smaller-scale feature maps provide more holistic depictions. When we develop DeepCrack on SegNet, we pairwise fuse the convolutional features generated by SegNet's encoder network and SegNet's decoder network. A single crack dataset is used to train and assess the DeepCrack network. DeepCrack surpasses existing state-of-the-art algorithms on all three tough datasets, achieving an F-Measure of more than 0.87 on average.