·Table of Contents
·Computer Processing and Simulation
Weld Defect Extraction and Classification in Radiographic Testing Based Artificial Neural NetworksNacereddine Nafaâ , Drai Redouane & BENCHAALA Amar
Laboratoire du Traitement du Signal et de l'Image Centre de Soudage et de Contrôle ( CSC ), Route de Dély brahim, BP 64, Chéraga, Algérie Tél/ Fax : (213) (2) 36 18 50 , E-mail : email@example.com
2.1. Ann's configuration
Because of digital radiogram's high resolution, it is physically impossible to develop a multilayer neural network that detects all the edges on radiogram at once.
Therefore, we designed a network window classifier which classifies the central pixel of a relatively small area in the image. To extract the edges, the window must slide, pixel by pixel, over the entire image. Then, the input layer of the network corresponds to the portion of image covered by the window (3x3) , . Therefore, the ANN has 9 neurons receiving the gray-scale values of pixels composing the square window. The output layer contains one neuron whom the state identifies the edges from the background and creates the segmented image.
There are no accurate rules for the option of the hidden layers number and the neurons number in each layer. Nevertheless, a hidden layer with 10 neurons has proven satisfying in the case of our application. The network synaptic weights are adjusted on the base of training couples set, representing examples that define possibilities of contours (rectilinear, circular, pointed, etc.). The used training rule is the back-propagation.Once, the system converges and the fixed total error is reached, the stage of training is finished and the network corresponding to the window is trained. Then, we apply it on all the image. The state of the output neuron identifies the contour. We has chosen 28 training couples representing cases of the most preponderant contours as shown in Fig. 1.
|Fig 1: Some cases of contours.||Fig 2: The neural network architecture|
|Fig 3: Three-layer neural network used to classify welded defects|
2.2 Applications and results
The proposed approach is tested on X-ray images used in industrial radiography. We can remark that for the X-ray image without noise as shown in Fig.4.a , the network extracts the different contours of weld defects, which are put in obviousness (See Fig.4.b). These defects are represented by dark stains with lengthened form that are positioned along the welded joint. Really, in the jargon of interpreters in radiography, it concerns a lack of penetration. The image-contours emerges very well the defect boundaries. This is not the case in Fig. 6.b . That is du essentially to the fact that the defect-image in Fig.4.a is better contrasted than the one in Fig. 6.a This is why, in this case, the application of contrast enhancement is recommended before the edge detection operation.
The obtained results show the effectiveness of using neural paradigms to detecting edges in X-ray images which are used in Non Destructive Testing. It has therefore succeeded to deliver directly contours of welding defects present in radiograms without the application of filtering techniques, contrarily to classic edge detection operators. Indeed, the proposed neural segmentation technique has provided satisfying results on noised or variable luminance images. However, the inconvenience of this technique resides in its speed of execution that is slow enough. This is why, in this type of applications, powerful and rapid computers are recommended. Currently, our work focuses on the validation of the proposed approach by other neural models such as Kohonen and Hopfield ANNs.
3.1Geometric invariant moments
Moments have been used as pattern features in a number of applications, to provide invariant recognition of 2D image patterns.
The regular moments mpq of a digital image pattern represented by f(x,y) are defined as :
p,q = 0,1,2,..
Hu introduced moments as image recognition features. Using non-linear combinations of normalized central moments, he derived a set of seven moments which has the desirable property of being invariant under image translation, scaling and rotation. Specifically, the central moments that have the property of translation invariance are given by :
The following normalized central moments are invariant under a scale change :
p+q = 2,3, . . .
The following moments Æ 1, Æ 2, . . . ,Æ 7 , are invariant under translation, rotation and scaling :
3.2. Neural network architecture
The conventional configuration for a multilayer neural network was used. This consists of an input layer of nodes with one node for each feature vector, a hidden layer, and an output layer with one node for each class. Each computational node (those in hidden and output layers) uses the sigmoid transfer function.The network weights were found by the back-propagation of errors technique.
Pattern vectors were generated by computing the invariant moments of the defects and then obtaining seven (7) values of moments (Æ'1, Æ'2, ... ,Æ'7) of each defect. The resulting seven moments were the inputs to the three layer feed-forward neural network shown in Fig.3.
The two neurons in the output layer correspond to the number of pattern classes (defect classes), and the number of neurons in the middle layer was heuristically specified as 10 . There are unknown rules for specifying the number of nodes in the internal layers of a neural network, so this number generally is based either on prior experience or simply chosen arbitrarily and then refined by testing.
3.3 Experimental results
We present results of combining the feature extraction technique by invariant moments and the neural network classifier described above using the welded defect X-ray images.
A. Invariant moment performance
In order to show the invariance performance of the proposed moments, we have used a set of test images representing a planer welded defect and a volumetric defect. Usual geometric transformations (rotation, scale change, the combination of the both and mirror effect) are applied on these images. Invariant moments are computed and compared between original images and their transforms. As shown in Figs and Histograms 7, the results are in reasonable agreement. The major cause of error can be attributed to the digital nature of the data.
B. Neural classifier performance
Initially, in he training process, the weights were initialized to small random values and the network was trained with the invariant moments corresponding to the planer and volumetric defects shown in Figs. 7. Each defect representation was assigned to a distinct output class. This set of couples (invariant moments, corresponding defects) are chosen as training set. This choice is justified by the fact that these defects can be considered as prototypes by referring to radiography interpreter suggestion. After the training phase was completed, the ability of the ANN to recognize defects was evaluated. The recognition phase was very simple for ANN and it consisted of only a feed-forward pass of the information presented to the input layer. The testing set is composed in the first part, by the transformed training defects by translation, rotation and scaling and in the second part by another set of defects as shown in Fig 8, and their transformed by the above transformations.
||Fig 7: Training set of the ANNs and invariant moment performance
a. Planer defect b. Volumetric defect |
|Type of Testing set||nb of tested defects||nb of classified defects|
|Transformed Training set||9||9|
|Non trained set||20||19|
|Transformed non trained set||75||72|
|Table 1: Accuracy of defect classification|
|Fig 8: Some tested defects from the testing set of the neural network|
In this paper, artificial neural network based approach for the classification of 2D dimensional welded defect images represented by translation, scale and rotation invariant region representation were introduced. ANN approach employing supervised learning represented by a multilayer ANN was utilized. The error back-propagation algorithm was used for the training of the multilayer ANN. Through experimentation with the defect-images for the classification problem, we show the feasibility of the proposed feature extraction and neural network paradigms, which are very promising in radiography inspection of welding joints. Presently, the use of a large bank of defects (porosity, leak of fusion, inclusions, etc.) as ANN training data is under investigation.
|© AIPnD , created by NDT.net|||Home| |Top||