Abstract:
DeepCrack is a continuously trainable deep convolutional neural network
designed for automatic crack detection and measurement. Cracks are linear
features that hold significant importance in various imaging applications, but
their detection is challenging due to factors such as poor road conditions, low
line quality, and low contrast. This work addresses these challenges by in troducing hyperconvolution techniques in DeepCrack. The proposed method
utilizes multi-scale depth convolution features learned through hyperconvo lution stages to effectively capture line structures. By combining a larger
feature map for a detailed perspective and a smaller feature map for a global
view, DeepCrack can accurately identify cracks. The network architecture of
DeepCrack is based on the encoder/decoder design of SegNet, allowing the
fusion of convolution functions from both networks on the same scale. To
evaluate the performance of DeepCrack, it was trained on one crack dataset
and assessed on three additional crack datasets. The morphological character istics of cracks, including length and breadth, were measured using a skeleton
extraction approach. The experimental results demonstrate the effectiveness
of the proposed technique, surpassing the performance of current state-of the-art methods. DeepCrack achieves an average F-Measure exceeding 0.90
on the challenging datasets, showcasing its superior performance. The pro posed technique demonstrates its potential for practical crack detection and
measurement tasks, accurately capturing different types of cracks, including
multiple cracks, thin cracks, and cross cracks. The findings highlight the suc cessful application of the suggested technique for crack measurement. The
superior performance of DeepCrack compared to existing methods under scores its significance in the field of crack detection
