Abstract: Aiming at the problems of weak defect features and complex background noise in digital radiography (DR) detection images of pressure vessel welds, traditional methods adopt gray histogram and global thresholding to implement defect segmentation and identification. As a single gray feature-based image processing method without integrating geometric and texture information, it suffers from low recognition accuracy and poor robustness. This paper presents an intelligent weld defect recognition approach based on the Gradient Boosting Tree (GBT) model. Without relying on complex deep networks, this method takes multi-dimensional features including gray, texture and geometry as inputs, and constructs a classification model suitable for weld defects through adaptive feature weight assignment and iterative fitting optimization. Meanwhile, a combined strategy of adaptive threshold segmentation, directional filtering and wavelet transform is adopted to achieve accurate extraction of defect features. Experimental results demonstrate that the proposed method, namely “gray + texture + geometry fusion + GBT model”, achieves 91.2% accuracy for crack defect recognition, 84.3% for porosity defect recognition, 91.0% for incomplete penetration defect recognition and 90.0% for lack of fusion defect recognition. The overall classification performance of the four key defects is significantly superior to that of traditional detection methods using only single gray feature. The model has the advantages of small parameters, fast inference speed and low hardware requirements, stable reliablility under small sample and complex working conditions, which is suitable for lightweight deployment in industrial fields. This method can effectively improve the detection efficiency and recognition accuracy of pressure vessel weld defects, and provide lightweight technical support for the intellectualization of industrial non-destructive testing.