This paper presents a methodology for extracting welds (linear or curved) from digitized radiographic images. The methodology consists of three major steps: feature extraction, pattern classification, and post-processing. Each weld image is processed line by line to extract three features for each object in the line image. These features are the width, the mean square error (MSE) between the object and its Gaussian, and the peak intensity (gray level). The fuzzy K-NN and fuzzy c-means algorithms are used as the pattern classifiers to recognize each object as weld or non-weld. Their performances are compared. The post-processing operation is applied to remove noises generated due to false alarms and to connect discontinuous weld lines due to misclassifications. The difficulties involved in post-processing results obtained by the fuzzy K-NN and fuzzy c-means algorithms are discussed. It is shown that both classifiers can successfully extract welds. However, the fuzzy K-NN classifier is concluded to be better because it gives fewer false alarms, and thus easier weld extraction.