Introduction
The welded joints radiogram often contain defects which the interpreter must identify and quantify, before he decides on their acceptability, by referring to non destructive testing standards and codes. Once the radiographic segmentation was accomplished providing a description in term of regions (defect and background), the problem is then to interpret their contents. It is thus question of determining effective attributes which permit to characterize these defect regions and to even recognize them like class elements easily identifiable. In industrial radiography, we can obtain radiograms on which weld defects, if they exist, can have various sizes and orientations.
For an example, a crack is identified as crack whatever its size and its orientation may be, and an inclusion is recognized as being an inclusion in spite of its position and its dimension. A major problem in the recognition of such defects is that these defects can be viewed from several angles and this, according to the orientation and the distance of the irradiated welded joint in regard to the radiation source. To characterize a given weld defect represented by its boundary or its region, the simplest attributes which be computed are the area and the perimeter . The latter cannot be used because of their sensitivity to geometric transformations.
For this reason, we will employ features which are invariant regardless geometric transformations of translation, rotation and scaling. A set of attributes satisfying the above conditions will be proposed in this paper. These geometric invariant attributes will follow from the calculation of geometric parameters (area, perimeter, etc.) and spatial moments. They will be implemented on binarized images issued from real radiographic films of welded joints. The main idea behind the principal component analysis (PCA) is to represent multidimensional data with less number of variables retaining main features of the data. It is inevitable that by reducing dimensionality some features of the data will be lost. It is hoped that these lost features are comparable with the “noise” and they do not tell much about underlying population.
The welded joints radiogram often contain defects which the interpreter must identify and quantify, before he decides on their acceptability, by referring to non destructive testing standards and codes. Once the radiographic segmentation was accomplished providing a description in term of regions (defect and background), the problem is then to interpret their contents. It is thus question of determining effective attributes which permit to characterize these defect regions and to even recognize them like class elements easily identifiable. In industrial radiography, we can obtain radiograms on which weld defects, if they exist, can have various sizes and orientations.
For an example, a crack is identified as crack whatever its size and its orientation may be, and an inclusion is recognized as being an inclusion in spite of its position and its dimension. A major problem in the recognition of such defects is that these defects can be viewed from several angles and this, according to the orientation and the distance of the irradiated welded joint in regard to the radiation source. To characterize a given weld defect represented by its boundary or its region, the simplest attributes which be computed are the area and the perimeter . The latter cannot be used because of their sensitivity to geometric transformations.
For this reason, we will employ features which are invariant regardless geometric transformations of translation, rotation and scaling. A set of attributes satisfying the above conditions will be proposed in this paper. These geometric invariant attributes will follow from the calculation of geometric parameters (area, perimeter, etc.) and spatial moments. They will be implemented on binarized images issued from real radiographic films of welded joints. The main idea behind the principal component analysis (PCA) is to represent multidimensional data with less number of variables retaining main features of the data. It is inevitable that by reducing dimensionality some features of the data will be lost. It is hoped that these lost features are comparable with the “noise” and they do not tell much about underlying population.
For this purpose, in this work, the principal component analysis technique will be used to reduce the number of the attribute variables. When the expert knowledge is not explicitly defined or cannot be represented in terms of statistically independent rules, artificial neural networks (ANN) may provide a better solution than expert systems, and they can efficiently learn nonlinear mappings through examples contained in a training set, and conduct complex decision making. Then, the ANN can be effectively updated to learn new features. In this project, a feed forward neural network trained by the backpropagation algorithm will be used for the weld defect classification task . This neuronal classification consists in assigning the usual types of weld defects met in practice to four categories according to their morphological characteristics. Other work was the subject of the use of ANN in the radiographic testing area. Authors in and use ANN in the weld defect segmentation in edges and regions respectively. ANNs were also used in the planer and volumetric weld defect classification using Hu’s invariant moments as features .
No comments:
Post a Comment