An alternative approach for measuring the mechanical properties of hybrid concrete through image processing and machine learning

Abstract

Image processing (IP), artificial neural network (ANN), and adaptive neuro-fuzzy inference system (ANFIS) are innovative techniques in computer science that have been widely used to predict the properties of materials in structural engineering. The stability of reinforced concrete structures mainly depends on the mechanical properties of concrete, i.e., its compressive strength fc and tensile strength ft. Different kinds of inexpensive cement replacement materials (CRM) can be used to form hybrid concrete (HC) with enhanced mechanical and other properties. In this study, the IP, ANN, and ANFIS methods are properly combined and used to predict the mechanical properties of hybrid concrete. For this, 162 cylindrical specimens of HC with 0%, 15%, and 25% silica fume and fly ash as replacement material with three mix ratios, 1:3:6, 1:2:4, and 1:1.5:3 were cast at 14 days and 28 days curing. The specimens were divided into three equal sets as follows: (i) the first to find the compressive strength, (ii) the second to find the split cylinder strength, and (iii) the third to develop a database of images. For the image acquisition, each cylinder of the third set is cut into three slices using a stone cutting saw, resulting in six faces and a total of 324 images (6 x 54). Photos (digital images) are then taken in fully controlled lighting conditions from a height of 600 mm between the concrete surface and the camera lens. The acquired images are pre-processed (converted to grayscale, cropped, and resized to 256 x 256 pixels), and the statistical features are extracted to predict the fc and ft by using ANN and ANFIS techniques. Finally, the predicted values are tested and validated through nondestructive testing methods. The actual values of the compressive and the tensile strength of concrete were compared to the corresponding values estimated by the proposed methods, i.e. (IP, ANN) and (IP, ANFIS). The accuracy of the results largely depends on the data set. The accuracy obtained by IP/ANN is 99.7% while the one obtained with IP/ANFIS is 97.8%.