The presented work addresses the problem of particle detection with neural networks in defocusing particle tracking velocimetry. A novel approach based on synthetic training data refinement is introduced, with the scope of revising the well documented performance gap of synthetically trained neural networks, applied to experimental recordings. In particular, synthetic particle image data is enriched with image features from the experimental recordings by means of deep learning through an unsupervised image-to-image translation. It is demonstrated that this refined synthetic training data enables the neural-network-based particle detection for a simultaneous increase in detection rate and reduction in the rate of false positives, beyond the capability of conventional detection algorithms. The potential for an increased accuracy in particle detection is revealed with neural networks that utilise small scale image features, which further underlines the importance of representative training data. In addition, it is demonstrated that neural networks are able to resolve overlapping particle images with a higher reliability and accuracy in comparison to conventional algorithms, suggesting the possibility of an increased seeding density in real experiments. A further finding is the robustness of neural networks to inhomogeneous background illumination and aberration of the images, which opens up DPTV for a wider range of possible applications. The successful application of synthetic training-data refinement advances the neural-network-based particle detection towards real world applicability and suggests the potential of a further performance gain from more suitable training data.