Studies on Augmented Data and Time Performance to Approach Crack Detection in Concrete

While Convolutional Neural Networks are generally considered efficient structures for image processing, that task has been shown to be non-trivial in cases of datasets of imbalanced and low-quality images. Here, it is not only significant to train an algorithm towards a high accuracy enough, but also to do that in a limited amount of time. This can be understood from that low-quality datasets may require several experiments on representations of hyperparameters for an efficient network. Such experiments may be costly, not only generally in terms of time, but also in terms of payments for cloud services as a development platform. This contribution builds upon previous investigations that showed promising techniques for attacking a problem with images of cracks in bridge concrete. Results from that work is here further improved through data augmentation techniques where simply inverting images is proven to improve the performance of the network. Besides this, results will be presented regarding a prognostication of the time required to train the proposed Convolutional Neural Networks.