Please use this identifier to cite or link to this item:
https://scholarbank.nus.edu.sg/handle/10635/222709
Title: | DETECTING THE PRESENCE AND LOCATION OF SURFACE DEFECTS FROM IMAGES | Authors: | TANG CHIN SING | Keywords: | Building PFM Project and Facilities Management Clayton Miller 2019/2020 PFM Computer Vision Machine Learning Convolutional Neural Network (CNN) Classifier Model Image Augmentation Training Accuracy Testing Accuracy |
Issue Date: | 2019 | Citation: | TANG CHIN SING (2019). DETECTING THE PRESENCE AND LOCATION OF SURFACE DEFECTS FROM IMAGES. ScholarBank@NUS Repository. | Abstract: | With advancement in technology allowing for cheaper and faster visual processors, various industries are racing to use of computer vision in the attempt to automate their work processes that would have otherwise depended heavily on the slower, less consistent and likely more expensive human visual system. While many works of literature reveal that such technologies are employed in different fields, the research is still limited on how it can be adopted in Singapore’s, or any, building and construction sectors. In line with the smart nation initiatives in Singapore, this report explores how the use of computer vision could help automate the traditionally manual defect inspection process, and potentially bring about greater efficiency, and effectiveness in the building industry. This dissertation will be a continuation of the study done by Chan (2019), where he introduced the use of a simple image classifier model to classify wet and dry building defects. In this dissertation, the aim of the study is to replicate the previous model, while exploring other training and Convolutional Neural Network (CNN) parameters, as well as other image augmentations, specifically blur, colour, and sharpness, to assess their impact on the model’s accuracy. The experiment results would help determine the ideal combination of training parameters to adopt for training the model as well as possibly disproving the general speculation that higher number of convolutional layers and image augmentations could improve the accuracy of the model. The results showed a 70% test accuracy vs a 50% test accuracy before and after the addition of the convolutional layer respectively. Also, the results of the all augmented datasets and the original dataset have test accuracies in the range of 60-70% showing no obvious sign that the augmentations have improved the model’s performance. With valuable insights gained on the classifier model, it is understood that the model adopted is rather over-simplified and the dataset to be of a relatively tiny scale, and thus the results obtained should not be generalised to all datasets as it is still highly dependent on many other factors. In the closing chapter, recommendations are proposed on how future studies could be improved. | URI: | https://scholarbank.nus.edu.sg/handle/10635/222709 |
Appears in Collections: | Bachelor's Theses |
Show full item record
Files in This Item:
File | Description | Size | Format | Access Settings | Version | |
---|---|---|---|---|---|---|
Tang Chin Sing 2019-2020.pdf | 4.59 MB | Adobe PDF | RESTRICTED | None | Log In |
Google ScholarTM
Check
Items in DSpace are protected by copyright, with all rights reserved, unless otherwise indicated.