Abstract: A normalized singular value decomposition (SVD) of image matrix was presented for the automatic recognition of welding targets.In order to eliminate the dependence of SVD on the orientation and scale of a target image, the direction and length of the major axis of the target image were first determined using the 2ed-order moments, and the x-coordinate of the target image was rotated to the direction of major axis and the length of major axis was normalized, and then the SVD of the target image was completed, and the singular value obtained by the normalized SVD method can be used as the feature matching of welding target recognintion because of its stability.The contrastive recognition experiments were made using the conventional direct SVD and proposed normalized SVD for a planar weldment image with different orientations and scales.In addition, the effect of the target image brightness on SVDwas also investigated. The experimental results shown that the normalized SVD method has high stability to the translation, rotation, scale and brightness of target images, and can be applied to the automatic recognition of welding targets.