To cater the need for real-time crack monitoring of infrastructural facilities,a CNN-regression model is proposed to directly estimate the crack properties from patches.RGB crack images and their corresponding masks o...To cater the need for real-time crack monitoring of infrastructural facilities,a CNN-regression model is proposed to directly estimate the crack properties from patches.RGB crack images and their corresponding masks obtained from a public dataset are cropped into patches of 256 square pixels that are classified with a pre-trained deep convolution neural network,the true positives are segmented,and crack properties are extracted using two different methods.The first method is primarily based on active contour models and level-set segmentation and the second method consists of the domain adaptation of a mathematical morphology-based method known as FIL-FINDER.A statistical test has been performed for the comparison of the stated methods and a database prepared with the more suitable method.An advanced convolution neural network-based multi-output regression model has been proposed which was trained with the prepared database and validated with the held-out dataset for the prediction of crack-length,crack-width,and width-uncertainty directly from input image patches.The pro-posed model has been tested on crack patches collected from different locations.Huber loss has been used to ensure the robustness of the proposed model selected from a set of 288 different variations of it.Additionally,an ablation study has been conducted on the top 3 models that demonstrated the influence of each network component on the pre-diction results.Finally,the best performing model HHc-X among the top 3 has been proposed that predicted crack properties which are in close agreement to the ground truths in the test data.展开更多
文摘To cater the need for real-time crack monitoring of infrastructural facilities,a CNN-regression model is proposed to directly estimate the crack properties from patches.RGB crack images and their corresponding masks obtained from a public dataset are cropped into patches of 256 square pixels that are classified with a pre-trained deep convolution neural network,the true positives are segmented,and crack properties are extracted using two different methods.The first method is primarily based on active contour models and level-set segmentation and the second method consists of the domain adaptation of a mathematical morphology-based method known as FIL-FINDER.A statistical test has been performed for the comparison of the stated methods and a database prepared with the more suitable method.An advanced convolution neural network-based multi-output regression model has been proposed which was trained with the prepared database and validated with the held-out dataset for the prediction of crack-length,crack-width,and width-uncertainty directly from input image patches.The pro-posed model has been tested on crack patches collected from different locations.Huber loss has been used to ensure the robustness of the proposed model selected from a set of 288 different variations of it.Additionally,an ablation study has been conducted on the top 3 models that demonstrated the influence of each network component on the pre-diction results.Finally,the best performing model HHc-X among the top 3 has been proposed that predicted crack properties which are in close agreement to the ground truths in the test data.
