Abstract
The advancement of smart cities has led to a growing demand for
efficient urban management systems, such as parking space management.
In this context, regularization has proven to be an essential
technique for improving the generalization of deep learning models.
This paper proposes a regularization method based on the use of smaller
data subsets across multiple epochs with resampling, aiming to
balance model learning and reduce the risk of overfitting. Using the
MobileNetV3-Small, four experiments were conducted with the PKLot
and CNRPark-EXT datasets. The results indicate that the proposed
method achieved competitive results while utilizing only a fraction of
the original data during training. By employing these strategies, it
was possible to reduce the data required per epoch by up to 97%, while
maintaining an average accuracy close to 89%. Furthermore, training
with the PKLot dataset highlighted the positive impact on model robustness
when using datasets with greater diversity and quantity. This
study underscores the importance of new regularization approaches
to enhance the efficiency and generalization of deep learning models
in urban applications.