The cracks in concrete buildings pose a direct threat to their stability and durability, as they allow aggressive agents to infiltrate, accelerating the degradation of the concrete and its mechanical properties and consequently compromising the structure's stability. A thorough analysis is essential to identify cracks that could weaken the entire structure and cause irreversible damage. With the constant expansion of concrete construction, the need for inspections continues to grow. However, manual inspections are time-consuming, subjective, unsuitable, and even impractical for traditional buildings in difficult regions. In recent years, due to the notable successes of artificial intelligence techniques in crack detection, many researchers have devoted themselves to developing various architectures and models to enhance the efficiency and performance of crack detection in concrete structures. In this context, this study presents the research methodology adopted to analyze the use of deep learning in automatic crack detection. It focuses on the convolutional neural network (CNN) process. A comparison is then made between conventional CNN architectures and those optimized by transfer learning based on various performance criteria. Additionally, the study highlights the use of CNNs for classification and other tasks, such as segmentation and feature extraction, to effectively detect cracks in concrete structures.