الفهرس 
INTRODUCTIONOnly 14 pages are availabe for public view
 |  | from | 267 |  |  |
| | | from | 267 | | |
Abstract
Crack inspection should be strictly carried out since it is closely related with the structural health and reliability especially in hazardous places where human can’t perform a certain task like working areas with electricity, possibility of fire, hazardous substances, overheating and electro-magnetic waves. However, it is difficult to find cracks by a visual check for the extremely large structures. So, the development of crack detecting systems has been a significant issue. Final objective of this research is to develop an automatic crack detection system that can analyze the surface and estimate the crack depth efficiently. The inspection should be done due to the problems that may occur like corrosion and deformation. The algorithm is composed of two parts; image processing and image characteristics. Neural network is used to learn the image characteristics. In this study, an automatic crack detection and depth estimation technique is developed using image processing and neural networks. Cracks are extracted sufficiently from the captured images using Crack ITv1 toolbox with measurement of maximum actual depths using Keyence (VK-X100) laser microscope and estimated depths using Make3D toolbox.Also the average intensities values for the extracted cracks are calculated per each one mm section.Feed forward back propagation and generalized regression neural networks are implemented and compared in order to get the best results for estimating the actual depth of cracks. The method is applied to eleven cracks with one hundred and fife segments. Eight cracks divided into one mm sections are used for training the neural network and three cracks are used for testing it. The reasonable results for the proposed neural networks structures are shown from generalized regression and feed forward back propagation. Both neural network structures show improvement in the detection of crack depth. While the feed forward back propagation (which is trained with estimated depths and extracted crack profile average intensities, output is measured depths) gave the most acceptable results with reduction in the overall error from 31.46% to 24.5%, the generalized regression (which is trained with estimated depths, output is measured depths) gave the best results with reduction in the overall error to 20.72%.