Recent studies have witnessed remarkable progress in harnessing convolutional neural networks (CNNs) to overcome the inherent limitations of conventional particle image velocimetry (PIV) methods. Traditional PIV techniques often suffer from compromised resolution and precision, hindering their ability to capture the complexities of fluid dynamics within the observation frame. While CNNs offer promise in addressing these challenges, they face obstacles such as limited accuracy, weak generalization, and a dearth of physical interpretability. In our prior research, we presented a CNN architecture that incorporates optical flow algorithms as supplementary physical constraints, thereby bolstering the model interpretability and precision. Nevertheless, the practical implications of this approach, especially when dealing with multi-dimensional, low-quality particle image data and restricted training sets, have yet to be fully explored. To address this knowledge gap, we have assembled a comprehensive dataset that simulates a wide array of experimental scenarios. We have systematically assessed the influence of velocity regularization on neural network performance, taking into account variations in image quality and the size of training datasets. The results underscore the pivotal importance of velocity regularization in enhancing the predictive prowess of neural networks, particularly when dealing with poor image quality and smaller data sizes. This study provides useful insights into the effective application of CNNs with velocity regularization in the field of experimental fluid dynamics.