At present, a number of computer vision-based crack detection techniques have been developed to efficiently inspect and manage a large number of structures. However, these techniques have not replaced visual inspection, as they have been developed under near-ideal conditions and not in an on-site environment. This article proposes an automated detection technique for crack morphology on concrete surface under an on-site environment based on convolutional neural networks (CNNs). A well-known CNN, AlexNet is trained for crack detection with images scraped from the Internet. The training set is divided into five classes involving cracks, intact surfaces, two types of similar patterns of cracks, and plants. A comparative study evaluates the successfulness of the detailed surface categorization. A probability map is developed using a softmax layer value to add robustness to sliding window detection and a parametric study was carried out to determine its threshold. The applicability of the proposed method is evaluated on images taken from the field and real-time video frames taken using an unmanned aerial vehicle. The evaluation results confirm the high adoptability of the proposed method for crack inspection in an on-site environment.

In many developed countries with a long history of urbanization, there is an increasing need for automated computer vision (CV)-based inspection to replace conventional labor-intensive visual inspection. This paper proposes a technique for the automated detection of multiple concrete damage based on a state-of-the-art deep learning framework, Mask R-CNN, developed for instance segmentation. The structure of Mask R-CNN, which consists of three stages (region proposal, classification, and segmentation) is optimized for multiple concrete damage detection. The optimized Mask R-CNN is trained with 765 concrete images including cracks, efflorescence, rebar exposure, and spalling. The performance of the trained Mask R-CNN is evaluated with 25 actual test images containing damage as well as environmental objects. Two types of metrics are proposed to measure localization and segmentation performance. On average, 90.41% precision and 90.81% recall are achieved for localization and 87.24% precision and 87.58% recall for segmentation, which indicates the excellent field applicability of the trained Mask R-CNN. This paper also qualitatively discusses the test results by explaining that the architecture of Mask R-CNN that is optimized for general object detection purposes, can be modified to detect long and slender shapes of cracks, rebar exposure, and efflorescence in further research.

Innovative concrete structure maintenance now requires automated computer vision inspection. Modern edge computing devices (ECDs), such as smartphones, can serve as sensing and computational platforms and can be integrated with deep learning models to detect on-site damage. Due to the fact that ECDs have limited processing power, model sizes should be reduced to improve efficiency. This study compared and analyzed the performance of five semantic segmentation models that can be used for damage detection. These models are categorized as lightweight (ENet, CGNet, ESNet) and heavyweight (DDRNet-Slim23, DeepLabV3+ (ResNet-50)), based on the number of model parameters. All five models were trained and tested on the concrete structure dataset considering four types of damage: cracks, efflorescence, rebar exposure, and spalling. Overall, based on the performance evaluation and computational cost, CGNet outperformed the other models and was considered effective for the on-site damage detection application of ECDs.

In most hyperspectral super-resolution (HSR) methods, which are techniques used to improve the resolution of hyperspectral images (HSIs), the HSI and the target RGB image are assumed to have identical fields of view. However, because implementing these identical fields of view is difficult in practical applications, in this paper, we propose a HSR method that is applicable when an HSI and a target RGB image have different spatial information. The proposed HSR method first creates a low-resolution RGB image from a given HSI. Next, a histogram matching is performed on a high-resolution RGB image and a low-resolution RGB image obtained from an HSI. Finally, the proposed method optimizes endmember abundance of the high-resolution HSI towards the histogram-matched high-resolution RGB image. The entire procedure is evaluated using an open HSI dataset, the Harvard dataset, by adding spatial mismatch to the dataset. The spatial mismatch is implemented by shear transformation and cutting off the upper and left sides of the target RGB image. The proposed method achieved a lower error rate across the entire dataset, confirming its capability for super-resolution using images that have different fields of view.