Computer vision-based surface crack identification technique using Gaussian process models

Abstract

The process of monitoring the health of civil, mechanical and aerospace structures, to ensure their continuous functioning, without any undesirable changes in their prescribed state, is popularly known as Structural Health Monitoring (SHM). SHM is being performed in two ways: global and local SHM. For global SHM, sensors are needed to be mounted in the critical areas to obtain the inputs such as excitation loads and outputs such as responses of the structure. This imposes difficulties such as cost, time and efforts to install, operate and maintain the instrumentation, cables, data acquisition and communication systems. The effective local SHM, which predominantly uses non-destructive testing methods such as X-ray, impact echo, ultrasonic methods, thermography, and Ground-Penetrating Radar (GPR), on the other hand, incur huge costs for the specialized equipment. Also, other issues such as inaccessibility of regions for instrumentation and stalling of structures during operation for monitoring are posing challenges in the existing global and local SHM approaches. Recently, with the low cost and effective imaging hardware combined with the robust computer-vision techniques, there is a greater shift towards the computer-vision based SHM, which can alleviate the challenges of other existing SHM approaches. In this paper, an effort is made to identify and classify the images into two classes (cracked surface and uncracked surface), using an improved version of Gaussian Process (GP) models in its regression form. The benchmark image datasets with and without cracks on the surfaces of concrete bridge decks, published by Utah State university, USA, are used for validation studies. Also, the paper compares and evaluates the performances of the various kernel functions used in the GP model during regression.