72 Y.-J. Cha and W. Choi Fig. 9.1 Overall architecture A possible breakthrough is the implementation of artificial neural networks [15]. Especially, convolutional neural networks (CNNs) have received a great attention in image classification due to the excellent performance [16]. In addition, extremely fast computations by the conjugation of graphic process units (GPUs) enable CNNs to have deeper architectures that can learn a vast amount of features from datasets [17]. In this study, a deep CNN architecture is proposed for detecting concrete cracks, and two GPUs (Nvidia Geforce Titan X 2ea) are used [18]. 9.2 Overall Architecture of the Proposed CNN and the Results The designed architecture is depicted in Fig. 9.1 representing how an input data is generalized by passing through the designed architecture. Overall, input images with 256 256 3 pixel resolutions are generalized to the vector with 96 elements in training, and the vector is consequently classified to crack or intact after rectified linear unit (ReLU), the last convolution, and softmax layers. During the training process, pooling layers simply pool features from previous layers. These operations may reduce computational cost significantly. Convolution layers conduct element by element multiplications between inputs of each layer and receptive fields, where the values of receptive fields are given by random numbers and tuned in stochastic gradient descent. The proposed CNN trained on 32K images and 8K extra images are used in validation, in which the images were cropped from raw images taken under uncontrolled conditions to generate a dataset with extensively varying features of cracks. As a result, the designed architecture records 98.25 and 98.04% of accuracies in training and validation respectfully. It is anticipated that the trained CNN classifier works very well in testing with images that are taken under uncontrolled situations because the generated training dataset includes extensively varying crack images, and the accuracies reach to about 98%. In the future, the trained CNN will be combined with a sliding window technique and autonomously flying drones. 9.3 Conclusion A vision-based method for detecting concrete cracks using a deep CNN was proposed. To generate the dataset with extensively varying crack features, images were taken under uncontrolled circumstances, and they cropped into small images of256 256. The generated image database was fed into the designed CNN. Consequently, 98% of accuracies in training and validation were recorded. In the future, we will combine the trained CNN with a sliding window technique and autonomously flying drones to develop an advanced framework.
RkJQdWJsaXNoZXIy MTMzNzEzMQ==