The detection of micro-defects within cemented carbides necessitates a high-frequency, high-sensitivity ultrasonic non-destructive testing transducer (UNDTT), whose performance is highly sensitive to geometric structural parameters. Conventional design approaches rely heavily on empirical trial-and-error, resulting in low efficiency and difficulty in achieving globally optimal solutions. To address this limitation, an intelligent multi-objective optimization method is proposed for transducer structural parameters—namely, radius, matching layer thickness, and backing layer thickness—to simultaneously maximize sensitivity ($V_{pp}$), center frequency ($f_c$), and bandwidth (BW). By investigating the relationship between structural parameters and performance metrics, a dataset was constructed and used to develop a convolutional neural network (CNN) surrogate model that captures their nonlinear mapping. The CNN was integrated with the NSGA-III multi-objective optimization algorithm to iteratively generate a Pareto-optimal solution set, from which the best design was selected using the entropy-weighted Technique for Order Preference by Similarity to an Ideal Solution (TOPSIS). Finite element analysis (FEA) validation confirmed prediction errors below 7.0%. Compared to conventional designs, the proposed approach delivers a 46.1% higher sensitivity and a 7.7% broader bandwidth while maintaining a thinner matching layer. These results confirm the effectiveness and practical advantage of the proposed framework. This data-driven approach offers an efficient alternative for designing a high-performance UNDTT.