Abstract: In order to realize the accurate identification of weld defects in large-scale thick-walled pressure vessels and to improve the accuracy of defect evaluation and detection efficiency, based on the mark-based improved watershed time flight of diffraction (TOFD) image segmentation and combined with the texture features of typical defect images, local phase quantization and local binary patterns are used respectively from the image spatial domain and frequency domain characteristics. The localized two value model can provide the local neighborhood characteristic parameters of the defect region, and through the normalization and fusion of the two feature parameters, the fusion feature vector is then classified by the support vector machine. The experimental results show that the fusion feature recognition rate proposed after detecting the 4 and 4 block of the image is the best, and the classification recognition accuracy rate reaches 87.1%.