Robust Deep Representation Learning For Road Surfaces Crack Detection

For Health and safety monitoring in civil constructions such as roads,bridges and culvets,identifying the region of interest is the fundamental requirement for image analysis at high-level semantic.One of the major structual problems in concrete and asphat structures is cracks which starts with harming the visual aspect of the construction and further lead to failure of the construction.Therefore,early identification of cracks is vital to maintain the service life of civil and transportation infrastuctures.Traditional visual inspection of road cracks which is ussually conducted through human visualization is expensive,time consuming and prone to errors as it depends on human judgements.The recent resurgence of Computer Vision(CV)based inspection has attracted a considerable attention and is progressively replacing traditional visual inspection which is normally conducted on-site,thereby handling considerable weakness and reducing challenges posed by traditional inspection methods.In view of that,the work in this dissertation proposes various efficient,robust and accurate CV algorithms for effective road cracks detection built on hybrid structures of Recurrent-Convolutional Neural Networks(RCNN),generative adversarial networks(GAN)and modified u-net architectures.These algorithms represents the main focus of this dissertation as explained in the following;The dissertation firstly proposes a robust deep representation learning algorithm for efficient road crack detection,benefiting from hybrid structures of multichannel parallel Convolutional Neural Networks(CNN).A unique hybrid framework is introduced,which utilizes low processing units to accurately perform image processing and analysis.Attention mechanism is further introduced allowing the model training to focus on small but important dataset with increased performance of the model.Bayesian Optimization Algorithm(BOA)were used to optimize the multichannel parallel Convolutional neural networks training with the fewest possible neural network layers to achieve maximum accuracy,improved efficiency and minimum processing time.Experimental results shows that,the proposed algorithm can achieve high accuracy around 95% in road surface cracks detection task which is good enough to replace traditional human inspection.Furthermore,this dissertation proposes RCNN-GAN which is an enhanced deep learning(DL)approach towards the detection of road crack.RCNN-GAN is deep learning based crack detection which combines two effective techniques RCNN and GAN with reduced layers to improve road cracks detection accuracy.RCNN is an object detection model that uses high-capacity CNNs to bottom-up region proposals to localize and segment objects.It deploy a selective search to detect the regional proposal network to detect object in any input image by defining boundaries to the Region of Interest(ROI)and then extracts features from each region independently for classification.GAN deploys unsupervised machine learning approach,that involves automatically discovering and learning the regularities or patterns in input data in such a way that the model can be used to generate new examples that plausibly could have been drawn from the original dataset.Through experiments,It has been found out that the combination of RCNN and GAN provides improved performance of the model.Additionally,this dissertation proposes a modified u-net architecture for road surface cracks detection,The architecture looks like a‘U’which justifies its name.This architecture consists of three sections: The contraction,The bottleneck,and the expansion section.The contraction section is made of many contraction blocks.Each block takes an input applies two 3×3 convolution layers followed by a 2×2 max pooling.The number of kernels or feature maps after each block doubles so that architecture can learn the complex structures effectively.The bottom most layer mediates between the contraction layer and the expansion layer.It uses two 3×3 CNN layers followed by 2×2 up convolution layer.The proposed u-net architecture detects cracks on the road surfaces by detection and classification of the road images thereby determining whether a particular image represents cracks or not.The work in this dissertation shows that computer vision algorithms can improve the performance of road surface cracks detection considerably with strong theoretical guarantees under complex dynamic patterns and variabilities in image datasets to meet the requirements of modern computer vision applications.Specifically,fundamental characteristic features such as effectiveness,scalability,robustness and efficiency against modern dataset and handling various image dataset.The dissertation demonstrates and validates empirically the effectiveness of the proposed computer vision algorithms via extensive experimental and rigorous evaluation on massive large-scale real-world image dataset.Experiments across different tasks and dataset show applicability,robust generalization,accurancy and superior performance of proposed DL frameworks compared to the well-known state-of-the-art methods in road crack detection tasks.