FEATURE EXTRACTION AND CLASSIFICATION OF RADIOGRAPHIC WELD IMAGES

Abstract

In the present scenario of industrialization, welding has become one of the prominent features for any development. Welding is defined as the process of joining the two materials so that the bonding exists between the materials. The welding process gives a permanent joint to the material which is joined together, but it also affects the properties of the constituents. Any kind of deficiency in the welding process gives rise to weld defects. The process of welding generates many flaws such as gas cavity, lack of penetration, porosity, slag, crack, lack of fusion, worm hole, and undercut. Welded material should be inspected accurately in order to ensure the quality of the design and operation. Reliable welding guarantees safety and reliability. The weld defects are captured by the traditional X-ray imaging. Interpretation of the radiographic weld image is a tedious task and also, it is difficult to distinguish and calibrate a large number of defects. Hence, there exists a need for detection of welding flaws so that the defects can be eradicated and serious damage is averted. Non – Destructive Inspection is one of the important aspects which are responsible for identifying the weld defects. It is widely used since it detects flaws without damaging the property of the objects. With the advent of computer technology, recent research exerts on finding a technological solution for accurate identification and classification of welding defects. The research for the development of an automatic or semi-automatic system for weld flaws (defects) classification using radiographic images of weld joints has grown considerably in the last decades. Image processing is one of the recent approaches which is being used for the identification and classification of weld defects. But, still there exists a need for an optimal algorithm that holds valid for the entire radiographic weld image database with better classification accuracy. In the present work, an algorithm has been designed to accurately classify weld defects. The first objective proposed was to develop a weld image database for multi flaws weld images. The second objective was to improve the quality of the radiographic weld images by using noise removal and other pre-processing techniques. Further, the images are segmented to get the information of the interest from the whole images, which was the third objective. The fourth and main objective was to extract the texture features using various basic texture feature extraction techniques, multi-resolution texture feature i extraction techniques and hybrid texture feature extraction techniques on segmented images for classifying them with the support vector machine (SVM) and artificial neural network (ANN) classifier into 9 different categories. For the purpose of development and validation of the algorithm, an image dataset is a primary requirement. In the present work, the image database has been created from Welding research laboratory, Department of Mechanical and Industrial Engineering, Indian Institute of Technology Roorkee, Roorkee. The images were radiographic in nature and were not in good quality. It is indeed a cumbersome process to analyze a radiographic image to identify the welding flaws. There were in all 79 radiographic images with 8 types of flaws and one without flaws. There were 08 images from gas cavity, lack of penetration consists of 20 images, 16 images have been captured from porosity, 16 images of slag, crack had 11 images, lack of fusion consists of 07 images, wormhole consists of 02 images, and undercut consists of 03 images. Also, 05 images were considered having no defect. The radiographic films obtained from the laboratory were scanned with a high-resolution scanner to convert it in images. Pre-processing is the first phase of image analysis. The purpose of pre-processing is to improve the quality of the image being processed. Since the images are acquired by the scanner for digitizing so they are in RGB format. Further, they are converted from RGB to gray scale in order to reduce the computational time. All the images obtained were of different sizes. The images were resized according to their region of interest for further processing. Radiographic images have generally salt and pepper noise, impulse noise, random noise etc. caused by natural interventions and sometime during scanning process electronics noise is usually added in these images. In the present work, the median filter has been used to remove the noise in radiographic weld image and it removes the noise very efficiently here without affecting any relevant information. Afterward, the contrast enhancement has been done by Contrast-limited adaptive histogram equalization method. It enhances the dynamic range of the image-pixel gray level and enhances the contrast. For Image Segmentation which is a key step in image processing for image analysis, and is essential for the extraction of image features especially for geometrical features. Hence, for getting better accuracy for classification, after pre-processing the images were segmented with various segmentation techniques such as gray threshold, ii edge detection, horizontal edge detection using an integral filter, contrast and horizontal response using an integral filter and multilevel thresholding. Thresholding is the simplest method of image segmentation. It was used for changing a gray scale image to binary images. Edge detection algorithms seek to detect and localize edges without any input or interference from humans. The proposed technique focuses to extract the texture features from the raw image and the segmented images to classify the database into 9 different categories of flaws. Feature extraction is a method of extracting information present in the images. It reduces the amount of resources required to define a large set of data accurately. Features have been extracted using gray level co-occurrence matrices, gray level run length matrices, local binary pattern, uniform local binary pattern, rotation invariant local binary pattern, rotation-invariant uniform local binary pattern, local binary pattern - histogram Fourier features, completed local binary pattern, adaptive local binary pattern, uniform adaptive local binary pattern, rotational invariant adaptive local binary pattern, rotational invariant uniform adaptive local binary pattern and binary Gabor pattern respectively. Feature extraction has been carried using full feature vector data of the above techniques and also reduced feature vector data. The LBP variants were used for the above purpose. In order to address the issue of image rotation effect, LBPri has been used. Also, LBPu2 is used to reduce the uniformity present in an image pattern. To overcome the disadvantages of the rotational invariant LBPri, the LBPriu2 is used. A detailed description of the result is presented in the chapter. Further to improve the classification accuracy of radiographic weld flaws DWT based feature extraction techniques have been proposed, where DWT decomposed sub-images have been processed with LBP variants and BGP texture feature extraction to get the tentative features. As DWT has the property to emphasize the directional information of the images, the texture features obtained from these directional sub-images further help to the enrichment of feature vector data, which in turn help in better discrimination of the radiographic weld Images. Amongst, the proposed DWT based texture feature extraction techniques, DWTBGP has yielded the best classification accuracy using ANN with 92.40% using 70/30 ratio of randomly divided database. The classification accuracy of flaws present in radiographic weld Images, hybrid texture feature extraction techniques has also been proposed where, segmented images iii have been processed with GLCM, LBP, LBPu2, LBPri, LBPriu2 and their respective combinations. The proposed hybrid local binary pattern variant texture feature extraction techniques have fetched better classification results. Eventually, the thesis is concluded with a summary of the work presented in the thesis and also focuses on the scope of future work. An attempt has been made in the thesis to classify the weld images accurately with relatively higher classification accuracy.