The comparations of different image classification models in facility management in the construction industry in the UK

Facility management and strategic asset management have become more and more important in construction project management. With the development in image classification under machine learning, more and more models have been created and updated towards the directions of increasing accuracy and/or efficiency. Since each model has their own advantages and disadvantages, it is vital to compare their prediction results on facility management in the construction industry. Thus, this study has collected image data from ten construction projects in the UK and used them as training data for three popular models in image classification, including ResNet152 v1, ResNet18 v1, and MobileNet1.0 modules. There are 11,526 collected images suitable for final training at the end. The collected images come with labelled 4 levels information and is enough for training. After the training, the comparations among all the three chosen models are made by using same predication images and comparing their prediction results. The comparations are made under different dimensions. The results show that the prediction results of all the three models are very similar. However, regarding the probability distribution among the top-5 options, the MobileNet1.0 and ResNet18 v1 models have more evenly distributed confidence level while the probability distribution of the ResNet152 v1 model is more concentrated on the top suggestion. The study helps to compare and evaluate the three image classification models in facility management in the construction industry. Thus, a better decision could be made when the construction companies want to choose a model for facility management in the future based on their different requirements and needs.