Efficient Labelling of Air Voids in Hardened Concrete for Neural Network Applications Using Fused Image Data

Every concrete structure incorporates a system of air pores of different sizes which is characterized by design parameters according to the respective building codes. The system’s characteristics affect properties of the material, specifically the durability of concrete exposed to freeze-thaw action. Testing techniques like the traverse line method, the point count method or the enhanced contrast method have been established and standardized to determine the relevant parameters according to construction material standards in order to empower the future life cycle analysis of the structure. These procedures are typically governed by a major effort in terms of specimen preparation and microscopic examination. We present an approach for a time-efficient procedure to analyze the air void system in optical microscopic images of hardened concrete by means of a neural network that is reliable and highly reproducible at the same time. Height-based segmentation of images using confocal laser scanning microscopy data allowed the composition of an image dataset with automatically labelled air voids. The set consists of 32,750 images of air voids with diameters between 25 µm and \(1.1\,\times \,10 ^3\) µm. By translating the gained information about the size and the location of air voids via a binary mask, we created a corresponding dataset generated by means of an optical microscope. The dense one-stage object detector RetinaNet with a Resnet backbone, fed with the optical microscopic image dataset, demonstrates the effectiveness of the method referring to localization and characterization of air voids in images of hardened concrete. The presented approach supports the successive characterization of the standardized parameters of the air void system and advances the modelling and prediction of structural durability with regard to freeze-thaw resistance.