A system for crack pattern detection, characterization and diagnosis in concrete structures by means of image processing and machine learning techniques

Abstract

A system that attempts to find cracks in a RGB picture of a concrete beam, measure the cracks angles and widths; and classify crack patterns in 3 pathologies has been designed and implemented in the MATLAB programming language. The system is divided in three parts: Crack Detection, Crack Clustering and Crack Pattern Classification. The Crack Detection algorithm attempts to detect pixels depicting cracks in a region of interest (ROI) and measure the crack angles and widths. The input ROI is segmented several times: First with an artificial Neural Network (NN) that classifies image patches in ¿Crack¿ or ¿Not Crack¿, then with the Canny Edge detector and finally with the local Mean and Standard deviation of the intensities. Then all neighborhoods in the mask are passed through special modified line kernels called ¿orientation kernels¿ designed to detect cracks and measure their angles; in order to obtain the width measurement, a line of pixels perpendicular to the crack is extracted and with an approximation of the intensity gradient of that line the width is measured. This algorithm outputs a mask the same size as the input picture with the measured angles and widths. The Crack Clustering algorithm groups up all the crack image patches recognized from the Crack Detection to approximate clusters that match the quantity of cracks in the image. To achieve this a special distance metric has been designed to group up aligned crack image patches; then with an algorithm based on the connectivity among the crack patches the clusters are obtained. The Crack Pattern Classification takes the mask outputs from the Crack Detection step as input for a Neural Network (NN) designed to classify crack patterns in concrete beams in 3 classes: Flexion, Shear and Corrosion-Bond cracks. The width and angles masks are first transformed into a Feature matrix to reduce the degrees of freedom of the input for the NN. To achieve a desirable classification in cases when more than 1 pathology is present, every angle and width mask is separated in as many Features matrices as clusters found with the Clustering algorithm; then separately classified with the NN designed. Several photos depicting concrete surfaces are presented as examples to check the accuracy of the width and angle measurements from the Crack Detection step. Other photos showing concrete beams with crack patterns are used to check the classification prowess of the Crack Pattern Classification step. The most important conclusion of this work is the transference of empirical knowledge from rehabilitation of structures to a machine learning model in order to diagnose the damage on an element. This opens possibilities for new lines of research to make a larger system with wider utilities, more pathologies and elements to classify.