Surface Classification of Damaged Concrete Using Deep Convolutional Neural Network

Image-based Detection of Surface Defects in Concrete during Construction

“…While multiple datasets of defects in concrete structures exist, they either focus on different defects or contain images scraped from the internet. Hung et al [6] published the most relevant dataset used for honeycomb detection, but this was only published after data augmentation, increasing relabeling effort. Two datasets are collected.…”

“…First, Metis Systems AG provided a set of honeycomb images. Second, similarly to Hung et al [6], a dataset was scraped from the internet, providing a baseline to similar scraped datasets from research and enabling comparison with a realistic dataset.…”

“…To the author's best knowledge, the following is the only work using regular camera images and applying machine learning (ML) for honeycomb detection and is used as a baseline for this thesis. Hung et al [6] showed that convolutional neural networks (CNNs) could classify concrete images with a precision of 93% and recall of 93% into honeycomb, crack, moss, blistering, and normal classes. When their dataset, concrete damage classification (CDC), was published, it was limited in size and was scraped from the internet, introducing a bias, as demonstrated in Section ??.…”

“…The second dataset was obtained using a google image search similarly to Hung et al [6], who collected images for four classes. However, since only the class of honeycombs is relevant for this work, only the keyword phrases honeycomb concrete and honeycomb on concrete surface were used, and the number of downloaded images was increased from 50 to 100 compared to Hung et al [6].…”

“…The second dataset was obtained using a google image search similarly to Hung et al [6], who collected images for four classes. However, since only the class of honeycombs is relevant for this work, only the keyword phrases honeycomb concrete and honeycomb on concrete surface were used, and the number of downloaded images was increased from 50 to 100 compared to Hung et al [6]. The first 100 images matching the keyword phrases were downloaded and images containing no honeycombs, those with watermarks obstructing the image, duplicates, and images of low quality were removed.…”