Crack detection is done by using invasive or non-invasive techniques. Invasive techniques, which usually involve surveying using specialized equipment, like infrared light, thermal testing, ultrasonic techniques, and testing concrete samples in the laboratory, are often time-consuming and complex processes. All the invasive and non-invasive methods require structural experts to analyze and interpret the available data . The findings of such methodologies are often subjected to human interpretation and knowledge. With advent of improved imaging capabilities and increased computational power, other non-invasive techniques using digital image analysis of concrete structures have gained a lot of momentum. In the last few decades, more than 50 articles discussed the problems of concrete crack identification using pre-processing and post-processing techniques. A comprehensive overview of such methods (including advantages and disadvantages) are presented by Mohan et al. .
More recently, deep learning-based methods using ANNs (artificial neural networks) and CNNs (convolutional neural networks) have been applied to automatically process the images for crack/failure identification in onshore concrete installations [5,6,7,8,9]. Several applications based on the use of CNN-type neural networks to identify surface concrete cracks have recently emerged with the development of artificial intelligence and deep learning technologies. Many of these methods could be characterized by high classification accuracy. Most of the published data spans the last decade. Kim et al.  presented a conference paper in 2011 in which they used a backpropagation neural network for 225 concrete images. The network was trained using 105 images, and the trained network was tested for 120 new images. The recognition rate of the crack image was 90% and non-crack image was 92%. Choudhary et al.  published a methodology for crack identification and detection using an object detection method. In their work, they have utilized over 205 256 × 256 resolution images, claiming a high crack detection accuracy reaching 96%. Ha  published their work in 2016 in which they have used an image segmentation method for automatic peak detection enabling concrete crack identification.
The first extensive data set of over 40,000 images of 256 × 256 resolution is used by Cha et al.  in their 2017 paper. They have applied a CNN-based deep learning method for concrete crack detection achieving very high accuracy of over 98%. A series of papers published in 2019 using deep learning methods for crack identification and detection. Papers published by Chen et al.  and Cao et al.  use CNN-based methods on the same set of 40,000 images of 227 × 227 resolution, both presenting good recognition accuracy of over 99% and 90%, respectively. The 2019 paper published by Lee et al. is distinguishable as they have used very high resolution (3120 × 4160) data set comprising 60,000 images for training and testing of their CNN-based crack identification method with a very high prediction accuracy of about 99%. A series of papers published in 2019 by Moon et al.  and Kim et al.  also use convolution neural networks.
More recently, Kim et al.  have used the same data set [6,13] and have further improved the accuracy of the CNN to 99.9%. Jitendra et al.  and Wenming et al.  have presented a comparison of different deep learning networks in their papers published in 2020. They have also used over 20,000 images of 1024 × 1024 higher resolution images. Long-short term memory (LSTM) based deep learning convolutional neural networks have also been applied for crack identification . This paper is unique in the sense that the use of the CNN-LSTM type method has not been presented before in literature for problems of crack identification.
A recent paper was published in 2021 where thermal image-based crack identification and detection is performed using a U-net type learning network . However, the prediction accuracy of the U-net based on thermal images is rather low and reaches only 78%. In 2021, a paper presented by Yang et al.  shows a comparison of three neural networks, Alexnet, VGGNet13, and ResNet18, to recognize and classify crack images. This paper also shows that the trained YOLOv3 model detects the crack area with a satisfactory accuracy.
Although all these methods claim very high accuracy, they often ignore the complexity of the image collection process itself. Almost all published papers deal with images captured in ideal laboratory-based conditions. None of the published papers have specifically considered the challenge of identifying concrete cracks in the presence of shadows. One possible approach for dealing with this problem is shadow detection and its removal. However, shadow removal is far from being a straightforward task. Early papers on shadow detection and removal have been presented by Finlayson et al. [19,20]. A comprehensive detail of the various shadow detection and removal methods is presented by Murali et al. . More recently, ANN-based deep learning methods have also been deployed for shadow detection and removal [22,23,24]. Although several researchers have tried to solve the problem of shadow removal; this task still remains a complex topic with moderate to good success . Shadow removal in concrete crack images often severely impacts the quality of the digital image, which makes further crack detection very challenging .
A new approach for realistic crack detection using the augmentation of existing crack data sets by complex shadow shapes is proposed in this paper. The presented methodology helps to automate the crack detection in real environmental settings. Moreover, it enables the use of the deep learning-based crack identification methodology for broader real-world applications, including the use of camera-carrying areal unmanned vehicles.
Sample images of concrete surfaces in the presence of shadows. Parts (a,b) show concrete surfaces without cracks. Part (c) shows concrete crack surface with crack.
Some sample images are used to demonstrate crack detection challenges in the presence of shadows using existing deep-learning-based methods. For this purpose, we will use the published data set of concrete crack images from Ozgenel , which consists of 40,000 images of surfaces with and without concrete cracks. A set of sample images is shown in Figure 2. This database of images has been widely used by researchers for the training and testing of multiple deep learning models for concrete crack detection.
Typical images after the application of shadow augmentation technique for images in Mendeley Concrete Crack Images for Classification data set: (a,b) depicts images without cracks; (c,d) depicts images with cracks.
There are still some practical challenges that could be envisaged in applying our method using an unmanned areal vehicle (UAV). Most important is to ensure the quality of images acquired by the UAV device. This challenge could be mitigated by using high-resolution imaging cameras, which are readily available these days. Another challenge could be the impact of concrete surface coloring or location of the concrete surface, e.g., onshore or offshore concrete structures (underwater concrete structures). Both of these challenges could be further mitigated with the help of image augmentation. Yet another challenge is with respect to the quantification of concrete crack size. The approach presented in this paper is geared toward concrete crack detection, which is the first step toward detection of concrete crack next step is the quantification of the crack, its size, dimensions, and type of crack. Whether it is a superficial crack, micro-crack, or failure crack. All of them come with their own challenges. Further development of our method is required to address the crack quantification questions. Approach presented in this paper is geared towards preventive maintenance and routine inspections to enable preventive maintenance to ensure that cracks do not lead to failure of the structure.
Although our proposed approach may appear simplistic, it is, however, the most particle approach as it does not require the creation of an entirely new deep learning network and nor does it require massive data collection exercises for images of concrete cracks in challenging conditions. Our approach also demonstrates that its possible to use image augmentation to improve the accuracy of existing established networks, which have been tested on a number of difficult problems.
M.P., contributed to manuscript preparation, developed shadow detection and removal codes, and developed crack classification models. P.P., contributed to manuscript preparation, image augmentation techniques, and preparation of classification image data sets. M.L., contributed to developing crack classification models used in the paper. U.O., contributed to the preparation of augmented images used in testing and training of the deep learning networks. I.T., contributed to manuscript preparation and preparation of augmented images used in testing and training of the deep learning networks. M.R., contributed to manuscript preparation, guidance of overall structure of the paper, and internal critical review and assessment of approaches presented in the paper. All authors have read and agreed to the published version of the manuscript.
You have now completed a major part of the 3D asset workflow for use in games or virtual reality software. In addition to texturing normal information on a model, you can texture with other material maps that create realistic skin, hair, or other materials from cracked concrete to scratched gun metal. For reference on this next workflow, do a web search on PBR materials using tools like Substance Painter* and Blender. 2b1af7f3a8