IEEE Transactions on Intelligent Transportation Systems · 2022 · 96 citations · 30 references
ComputingCivil EngineeringComputer ScienceComputer VisionMachine LearningPavement EngineeringDeep LearningEdge DetectionImage SegmentationInfrared Remote SensingObject DetectionPattern RecognitionBenchmark Dataset EvaluationConvolutional Neural Network DetectionConvolutional Neural Network SegmentationConvolutional Neural NetworksCrack DetectionCracksInfrared Crack DetectionInfrared Image Fusion
Two issues exist in the convolutional neural network (CNN) used for asphalt pavement crack detection: balance between accuracy and complexity, and indistinct edges of cracks and asphalt pavement surface. Infrared thermography (IRT) could use the temperature difference between cracks and pavement surface to better distinguish cracks. This work aims to propose a robust crack detection method based on CNN and IRT. An open benchmark dataset was built for crack detection based on three types of images, including visible images, infrared images, and the fusion of visible and infrared images (in short, fusion image). The dataset also considers different conditions and periods, including single, multi, thin, and thick cracks; clean, rough, light, and dark backgrounds, and three periods in a day. Seven CNN segmentation models are trained and evaluated on this dataset. To keep a balance between accuracy and complexity, evaluation metrics (accuracy, and computational and model complexity) are used to have an overall evaluation of models rather than only the accuracy. The results show that the accuracy and predictions of the visible image and fusion image are almost identical for all models, which are much better than that of the infrared image. When the background is rough or cracks are similar to the background, the fusion image is a better choice for crack detection. Compared with the visible image, all segmentation models have a more stable performance for the fusion image. Among segmentation models, Feature Pyramid Networks (FPN) could be the best model because of its high accuracy and low complexity.
30
Deep Residual Learning for Image Recognition
Kaiming He, Xiangyu Zhang, Shaoqing Ren et al. · 2016 · 222.1K citations · Full text
Adam: A Method for Stochastic Optimization
Diederik P. Kingma, Jimmy Ba · UvA-DARE (University of Amsterdam) · 2014 · 84.8K citations · Full text
Fully convolutional networks for semantic segmentation
Jonathan Long, Evan Shelhamer, Trevor Darrell · 2015 · 36.9K citations
Feature Pyramid Networks for Object Detection
Tsung-Yi Lin, Piotr Dollár, Ross Girshick et al. · 2017 · 28.7K citations