Classifying Crack Types on Masonry Surfaces with Artificial Intelligence & Image Processing

Bandara, Yasaswin

dc.contributor.author	Bandara, Yasaswin
dc.date.accessioned	2022-12-20T05:12:21Z
dc.date.available	2022-12-20T05:12:21Z
dc.date.issued	2022
dc.identifier.citation	Bandara, Yasaswin (2022) Classifying Crack Types on Masonry Surfaces with Artificial Intelligence & Image Processing. BEng. Dissertation, Informatics Institute of Technology	en_US
dc.identifier.issn	2018162
dc.identifier.uri	http://dlib.iit.ac.lk/xmlui/handle/123456789/1189
dc.description.sponsorship	"A large number of building structures are built based on masonry in the world. Due to natural causes, mediocre construction or external forces inevitably cracks will occur. At present, these cracks are observed and analysed manually. Since the results will depend on the expert's experience and human errors can affect the accuracy. Compared to solutions that are implemented for other surfaces using computer vision to classify crack types and analyse severity, less focus was given to masonry structures. This research classifies the masonry wall crack types and analyses the severity of that type from a mobile application. In addition, a dataset that is suitable to classify crack patterns formed by gathering from various locations along with certification of a civil engineer. The model trained on masonry image patches using the MobileNet algorithm with transfer learning. Weights trained on the ImageNet dataset were used in the model. Then the fine-tuning technique was utilized by training only the batch normalization and custom layers of the model while freezing the MobileNet layers. The last layer of the model is ignored and custom layers are added. Dropout layers of 0.5 were added before and after the dense layer. The dense layer was defined with a rectified linear activation function. Next, it’s passed to a Softmax activation layer for classification. And augmentation methods have made the model more accurate. Pre-processing techniques and enhancements were employed to make the prediction more accurate and quicker. Adamax optimizer was used along with sparse categorical cross-entropy as the loss function. Embedding the TensorFlow Lite model to the mobile has shown optimal performance as it runs locally. An accuracy of 85% was achieved for the model. Then respective F1 scores were achieved for flexural cracks at 87%, non-cracked at 88%, shear cracks at 84% and torsional cracks at 80%."	en_US
dc.language.iso	en	en_US
dc.subject	Masonry Crack Detection	en_US
dc.subject	Computer Vision	en_US
dc.subject	Transfer Learning	en_US
dc.subject	Mobile Detection	en_US
dc.title	Classifying Crack Types on Masonry Surfaces with Artificial Intelligence & Image Processing	en_US
dc.type	Thesis	en_US