| J. Molleda,
"Computer vision techniques for real-time 3D shape inspection of rolled products",
University of Oviedo (Spain), 2008. (2008-07-17) Supervised by: D.F. García and R. Usamentiaga. |
|
| Abstract: | |
This thesis presents the application of computer vision techniques in the analysis, design and implementation of a 3D surface reconstruction system with real time constraints. It is focused on automatic quality inspection of rolled products in the steel industry. Quality evaluation of rolled products, so-called steel strips, requires apart from other analysis their flatness measurement. This measurement can be obtained by means of a three-dimensional reconstruction of the strip shape.
Flatness can be defined as the degree to which the surface of a flat product approaches a plane. Wrong rolling process generates internal strains in the steel which results in flatness defects in the rolled strip. A few years ago, flatness measurement was carried out manually by human operators of the industrial installation. In order to manually measure the strip flatness, its advance on the installation is stopped, then the length of a specific surface interval and the difference between its shortest and highest level are measured. This measurement obtains the length and the extent of wave of the strip surface. This task is extremely tedious and time-consuming, therefore, only a small number of measurements for each strip can be obtained. From these measurements the type and magnitude of the flatness defects can be inferred. Growing demand in finished quality of rolled products and current demanding productivity requirements in the steel industry make manual inspection of any product not feasible.
Automatic flatness measurement of a steel strip is based on acquiring several length measurements of its surface and comparing them. If the strip is flat, all length measurements will be equals. Otherwise, differences between these length values will be observed as flatness defects. Automation of flatness measurement process in industrial environments is a recent topic. There are two flatness measurement system types according to their basis principle: mechanical and optical systems.
The reported work presents an overview about flatness and its measurement over rolled products, as well as a review of the more significant flatness measurement industrial systems. Moreover, computer vision techniques applicable to 3D surface reconstruction are described. Also, the design of a flatness measurement system of steel strips based on laser triangulation technique is presented.
Due to the great difficulty of generating appropriate synthetic data and the impossibility of repeating the industrial production process, several real time image compression algorithms have been evaluated and a real time grabbing service has been developed. The service allows building a data base which represents the wide range of surface conditions of a steel strip as well as the illumination conditions in its processing stage.
Existing techniques are mainly focused on the accuracy of laser detection and they are not robust, so their application on industrial environments is not possible. In this thesis an accurate and robust method to extract the laser profile on adverse conditions is proposed. The proposed method is divided in two phases. Firstly, a local search process of the laser out of the strip surface is carried out. Then, based on the knowledge obtained from previous phase, a global search of the laser on the strip surface is carried out. This second phase consists of an adjustment of curve segments based on a split and merge technique.
Flatness measurement system validation is carried out indirectly by means of the accuracy obtained in the reconstruction process of the three-dimensional shape of the strips. Results presented on this thesis have been obtained form real data from the steel industry.