The Effectiveness of Data Augmentation in Construction Site-Related Image Classification

Abstract

In modern construction, the construction sites are congested, busy, and full of obstacles. The environment in the construction site is dynamic. In recent times, deep learning-based models remain the main tools used for image classification. However, the performance of the deep learning models in construction site-related image classification is not convincing due to the dynamic nature and busyness of the construction sites. The availability of construction site-related image datasets also remains an obstacle in achieving the best performance from the deep learning models. Data augmentation is a technique used to apply random but realistic transformation to the images. Data augmentation will not only help to diversify the image dataset but also assist in increasing the size of the dataset. This study used a state-of-the-art YOLOv4 deep learning model and implemented data augmentation techniques like gamma transformation to control the intensity of light and mimic sunny, cloudy, day, and night situations in the construction site images. The other data augmentation technique used is Gaussian blur to minimize the details in the images, and salt-and-pepper noise to degrade the quality of construction site images. The model is trained and tested on Alberta Construction Image Dataset (ACID) and construction workers hand signal image datasets, with and without the implementation of data augmentation. The performance of the model is evaluated based on the test dataset while keeping all the parameters of the model the same. It is observed that the model trained on the augmented dataset performed better than the model trained on the non-augmented dataset by 4%.