We present and compare two different approaches for NDT multi-sensor data fusion at signal (low) and decision (high) levels. Signal-level fusion is achieved by applying simple algebraic rules to strategically post-processed images. This is done in the original domain or in the domain of a suitable signal transform. The importance of signal normalization for low-level fusion applications is emphasized in regard to heterogeneous NDT data sets. For fusion at decision level, we develop a procedure based on assembling joint kernel density estimation (KDE). The procedure involves calculating KDEs for individual sensor detections and aggregating them by applying certain combination rules. The underlying idea is that if the detections from more than one sensor fall spatially close to one another, they are likely to result from the presence of a defect. On the other hand, single-senor detections are more likely to be structural noise or false alarm indications. To this end, we design the KDE combination rules such that it prevents single-sensor domination and allows data-driven scaling to account for the influence of individual sensors. We apply both fusion rules to a three-sensor dataset consisting in ET, MFL/GMR and TT data collected on a specimen with built-in surface discontinuities. The performance of the fusion rules in defect detection is quantitatively evaluated and compared against those of the individual sensors. Both classes of data fusion rules result in a fused image of fewer false alarms and thus improved defect detection. Finally, we discuss the advantages and disadvantages of low-level and high-level NDT data fusion with reference to our experimental results.