Abstract & Overview
Although deep neural networks(DNNs) have become a key driver of visual scene understanding, a shortage of construction training images holds DNNs back from achieving its maximum performance potential.To unlock its full potential, this research proposes an automated virtual image synthetization framework and examines the effectiveness of synthetic images as DNN training data.This research particularly focuses on a common real-world application scenario wherein there is only a small, biased real training image dataset available.To this end, this research (1) builds a large amount of synthetic image dataset using the proposed framework; (2) integrates the synthetized images into a small, biased real image dataset; and (3) trains a DNN model with the integrated dataset and evaluates its performance in diverse field conditions.Results show that the integrated dataset-trained model always yields a higher performance than the one trained with only a small, biased real dataset.This finding suggests that synthetic construction images have encouraging potential in enhancing a given real training image dataset's quantity and visual scene diversity.It is also observed that the integration of real and synthetic images can improve the DNN performance by 51.2% while reducing the number of real training instances by 75%(from 2,000 to 500).This finding is more promising granted that the proposed framework can automatically generate and label synthetic construction images without manual inputs.This research can eventually help address the shortage of construction training images and develop a more accurate and scalable DNN model for visual scene understanding.