To address the issues of inconsistent results and strong subjectivity in the evaluation of multi-modal detection for small-diameter thin-walled tubes, a multi-modal fusion recognition method incorporating a physical similarity matrix and a dynamic weighting mechanism is proposed, and information fusion is achieved based on D-S evidence theory. Taking phased array ultrasonic testing (PAUT) and radiographic testing (RT) as the objects, a multi-feature discrimination model is constructed, and dynamic weight allocation is carried out in combination with the physical similarity matrix to correct and optimize the fusion of evidence conflicts. In this paper, three groups of Φ60 mm × 5 mm specimens were used for verification, and no misjudgment occurred in 10 groups of engineering samples. Under the condition of 900 groups of extended samples constructed based on real statistical distribution, the recognition accuracy of the method proposed in this paper reached 94.6%, and the AUC was 0.977. Compared with the traditional D-S, Yager and Murphy methods, the proposed method has higher recognition accuracy and lower conflict rate. This method improves the consistency and reliability of multi-modal detection results.