Flaws classification using ANN for radiographic weld images

This paper illustrates a novel approach for weld flaw classification incorporating texture feature extraction techniques and measurement of geometrical feature using Artificial Neural Network (ANN) classifier. The radiographic films of weld have been digitized first using digital camera, then these images are converted to gray image and region of interest are selected to reduce the processing time. Noise reduction and contrast enhancement techniques were implemented to assist in the recognition of weld region to identify the weld flaws. Further various segmentation techniques like edge base, region growing and watershed have been applied and tested on images to choose the best one for each flaws. Each of the delineation techniques are not equally important and worth for all types of flaws. Subsequently a different set of texture feature based on gray level co-occurrence matrix (GLCM) and measurement of geometrical features which characterize the flaws shape is extracted for each segmented image and given input to cascade-forward back propagation neural network using Levenberg-Marquardt training function. The classifier is trained to classify each of the image into different flaws categories. The proposed system delivers an overall classification accuracy of 87.34% for radiographic images of nine different types of weld flaws.