• Login
    View Item 
    •   Mak IR Home
    • College of Engineering, Design, Art and Technology (CEDAT)
    • School of Built Environment (SBE)
    • School of Built Environment (SBE) Collections
    • View Item
    •   Mak IR Home
    • College of Engineering, Design, Art and Technology (CEDAT)
    • School of Built Environment (SBE)
    • School of Built Environment (SBE) Collections
    • View Item
    JavaScript is disabled for your browser. Some features of this site may not work without it.

    Image- based automatic crack edge detection and extraction using deep neural network.

    Thumbnail
    View/Open
    LILLIAN'S DISSERTATION REPORT.pdf (2.292Mb)
    Date
    2020-10-08
    Author
    Nakibuule, Lillian
    Metadata
    Show full item record
    Abstract
    Timely recognition and repair of crack edges is of great significance for overall structural health monitoring and prevention of further damages. Currently, image processing techniques are being used to manipulate images in order to extract crack features however, due to extensively varying real world situations such as lighting and shadow conditions, the adoption of these techniques becomes challenging. There is need for a robust crack edge detection and extraction method with excellent capabilities of mining useful information from different image conditions that can be used for diagnosis and selection of appropriate rehabilitation methods to fix the damaged building elements. This study is centered on deep learning techniques to develop a crack edge detection and extraction model based on python programming language and responses given in JavaScript Object Notation (JSON).The model was trained on 154 images of 256*256 px validated on 22 images of 256*256px and finally tested on 44 images of 4000*3000 px which did not take part in the training and validation sets. A well generalized model was hosted on an open Application Programming Interface (Nanonets API) server .From the results the model scored an average accuracy 97.75%, these results are comparable to those from traditional filter-based edge detection algorithms like spatial domain Laplacian of Gaussian (LoG) and frequency domain Butterworth under different image conditions which registered average accuracies of 88.25% and 83.75% respectively. Furthermore the CNN based method showed a very robust performance in identifying the minimum crack width deviating by 4% from the Euro code international standard (=<0.5mm) and the traditional methods by 8.1% for Laplacian of Gaussian and 20% for Butterworth. The resultant crack information was modeled and stored in a flexible and interactive file Geodatabase.
    URI
    http://hdl.handle.net/10570/8356
    Collections
    • School of Built Environment (SBE) Collections

    DSpace 5.8 copyright © Makerere University 
    Contact Us | Send Feedback
    Theme by 
    Atmire NV
     

     

    Browse

    All of Mak IRCommunities & CollectionsTitlesAuthorsBy AdvisorBy Issue DateSubjectsBy TypeThis CollectionTitlesAuthorsBy AdvisorBy Issue DateSubjectsBy Type

    My Account

    LoginRegister

    Statistics

    Most Popular ItemsStatistics by CountryMost Popular Authors

    DSpace 5.8 copyright © Makerere University 
    Contact Us | Send Feedback
    Theme by 
    Atmire NV