© 2022, TUBITAK. All rights reserved.This study is about autonomous crack detection in concrete structures using computer vision and deep learning. Fast and successful detection of cracks is very important for early and effective detection of structural damage. In this context, 3 different data sets obtained from the online sources with crack photos and a data set with photos of different parts of different structures without crack photos were used. The results obtained by using different combinations of datasets for training and testing of deep learning architecture are discussed. U-Net, designed for image segmentation, was used as a deep learning architecture. The findings showed that the success rate of crack detection increased with the use of crack-free photographs of different parts of the structures in the training data in convolutional neural networks trained for crack detection in reinforced concrete structures. In addition, the results obtained by using different data sets for the training and testing of the convolutional neural network have given a serious insight that the U-Net architecture can be used in real-world crack detection problems. It is hoped that this study will be useful for researchers who want to research or develop projects on structural crack detection and introduce the points that researchers can benefit from.