GRAY LEVEL CO-OCCURRENCE MATRIX FEATURE EXTRACTION ANDPROBABILISTIC NEURAL NETWORK FOR WELDING DEFECTS RECOGNITION. In thispaper, we have developed Defects Recognition Systems on Metal Welding based on Gray Level CooccurrenceMatrix Texture Feature. The X-ray films used in this research are IIW (International Instituteof Welding) standard of radiograph technique. The first step is X-ray film digitalized, using a digitalcamera on X-ray interpreter. The next step is feature extraction; in this step co-occurrence matrix isdeveloped. The value of angular second moment, correlation, inverse difference moment and entropy arecalculated as texture feature on one distance and four directions. Probabilistic Neural Network is used as adefect classifier to classify the output of the systems. Recognition output is classified into 8 classes, thatis 1st class (normal/without defect), 2nd class (distributed porosity), 3rd class (incomplete penetration), 4thclass (burn through), 5th class (cluster porosity), 6th class (excessive cap), 7th class (excessive penetration) and 8th class (incomplete fusion). Three methods of training testing data sampling (random, semi randomand choosing) are compared in this research. Average of recognition on this system is 99,54 % using training testing paradigm 1 : 1.Key words: Welding defect, Recognition, Neural Network, Texture Extraction.Adhi Harmoko S*, Benyamin Kusumoputro**, Makmur Rangkuti***; * Departemen Fisika FMIPA, Universitas Indonesia, Depok 16424, email : adhi_hs@fisika.ui.ac.id ** Fakultas Ilmu Komputer, Universitas Indonesia, Depok 16424 *** PUSDIKLAT BATAN, Ps Jum’at, Jakarta Selatan
