**1. Introduction**

Much of the important concrete structures that are in use today were erected several years ago and now they are close to their design life [1]. As a result, the structures require regular condition assessment for understanding of the current state of their structural components. Condition assessment of a concrete structural element involves monitoring of displacements and evaluation of crack development visible on the surface of the element during laboratory mechanical tests. Monitoring is often done using wired contact sensors such as linear-variable-differential transformers (LVDTs). However, these sensors are often difficult to install and maintain.

Crack evaluation typically involves visual inspection by trained staff and simple measuring tools such as a Brinell magnifier. However, such inspection methods can be expensive, dangerous and time-consuming. On the other hand, computer vision methods allow for fully automated extraction of important information from digital images [2,3]. As a consequence, various methods leveraging image-processing techniques and computer vision methods such as digital image correlation and convolutional neural networks have been developed and applied in past decades [4–9].

Monitoring and measurement of displacements using computer vision methods are often performed by applying optical flow-based algorithms such as digital image correlation [10]. A detailed review of DIC applications can be found in [7,11]. DIC have been applied for measuring displacements and strains of specimens made of various materials. Mróz et al. [12] presented a feasibility study of DIC in determining strains in concrete exposed to fire.

DIC methods have also been used for assessment of concrete crack development. Helm in [13] shown how to use DIC for assessment of specimens with multiple growing cracks. Similarly, Rui et al. in [14] presented DIC-based measurement system of crack generation and evolution during static testing of concrete sleepers. In [15] Gehri et al. shown a study on an automated crack detection and measurement based on DIC. Finally, DIC techniques allow for measurements and calculations of strains localization and the width of the fracture process zones on the surface of notched concrete beams [16,17].

Crack assessment using DIC methods is very precise but also require huge computational resources and is very time-consuming. As a result, it is mainly used off-line for assessment after the tests. It is also possible to apply other non-destructive monitoring techniques such as Acoustic Emission (AE) [18] or microwave sensors [19]. On the other hand, in recent years, convolutional neural networks (CNN) have been developed and applied for online automatic detection of concrete cracks and structural damage. See for example, recent state-of-the-art reviews [5,6]. In [20] Cha et al. described an autonomous system for structural visual inspection using region-based deep learning for detecting multiple damage types. In [21] a system for real-time crack assessment with wall-climbing unmanned aerial system is presented. Roberts et al. in [22] shown a system for low-cost pavement condition health monitoring and analysis. In [23] Deng et al. presented a region-based CNN with deformable modules for visually classifying concrete crack. Finally, an application of CNN for detection of flaws in concrete using ultrasonic tomography is described in [24].

These two approaches to monitoring and evaluation of surface cracks in concrete structural elements can be combined for better description and assessment of concrete elements. To the best knowledge of the authors of the paper, there is no such a study on combining DIC and CNN algorithms in this context. As a result, this paper is organized as follows. In this first section, the motivation and main goals of the research study undertaken are given. In Section 2, a detailed description of the proposed new method is presented. In Section 3, CNN model development and deployment is outlined. In Section 4, the experimental procedure is described. In Section 5, the results and discussion are given. Finally, in Section 6, the most relevant conclusions are drawn.
