DEEPCRACK: LEARNING HYPERCONVOLUTIONAL FEATURES FOR AUTOMATIC CRACK DETECTION AND MEASUREMENT

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