Welding defect detection from radiography images with a cepstral approach

This paper presents a new approach for feature extraction from radiography images acquired with gamma rays in order to detect weld defects. In this approach, images are lexicographically ordered into 1D signals. Then, Mel-Frequency Cepstral Coefficients (MFCCs) and polynomial coefficients are extracted from these signals, one of their transforms, or both of them. Discrete Wavelet Transform (DWT), Discrete Cosine Transform (DCT), and Discrete Sine Transform (DST) are tested and compared for efficient feature extraction. Neural networks are used for feature matching in the proposed approach. Sixteen radiography images containing seventy three weld defects are used to evaluate the performance of the proposed approach. For performance evaluation, the tested images are degraded by Gaussian, impulsive, speckle, or Poisson noises with and without blurring. The experimental results show that the proposed approach can be used in a reliable way for automatic defect detection from radiography images in the presence of noise and blurring.