DETECTING BUILDING DEFECTS USING AN IMAGE CLASSIFIER MODEL

Abstract

Although computer vision has been around since the early 1960s, it has only been starting to gain traction in the recent years. Many industries already possess the technology and knowledge to apply computer vision to automate vision-based activities, improving productivity and efficiency, while reducing their consumption of resources. However, this is still not the case in the built environment in Singapore. As a national goal to transform Singapore to become a Smart Nation, use of smart technologies must penetrate all industries, this broadly includes using computer vision solutions in the built environment. However, with hardly any computer vision solutions implemented, and even less computer vision research published in the built environment, the study undertakes an experimental approach into using a unique computer vision solution to overcome problems in the built industry. This refers to the use of a proof-of-concept image classification model in detecting building defects.

Image data was collected from several of Singapore’s public housing estates and these images were fed into a custom-built image classification model to train the model to recognize between wet and dry defects. Since the initial selection for algorithms to programme this proof-of-concept model was done on a ‘trial and error’ basis, it took months before the study arrived at the decision of using the MaWario Image Classification Model. With a training image dataset of 800 images, consisting of 400 wet defect images and 400 dry defect images, the images were used to fit the lightweight and computationally cheap proof-of-concept model and tested against a 200 image testing dataset. The model was also used to identify image augmentation parameters to improve the accuracy of the test. Overall, the proof-of-concept model came back with optimistic results, with testing accuracy hovering around 80%, despite its shallow architectural layers.

In order for this proof-of-concept model to be used in the real-world, further research and experimentation must be conducted to improve the overall accuracy of the model such that it is able to reach accuracies of at least 98% or more. Thus, recommendations for future works would include an increase in allocation of training time and the use of a larger dataset. Identifying the more relevant image augmentation parameters and testing ii them concurrently can also help to determine the best model training for improved versions of this proof-of-concept model.