Vehicle Type Recognition using an Efficient Regularization in Mask_RCNN

A Vehicle Type Recognition (VTR) system faces challenges in achieving accurate classification when distinguishing between vehicle types with intra-class patterns, such as sedan cars, taxis, vans, minivans, trucks, and buses. The main challenge lies in effectively extracting and preserving discriminant features for each vehicle type to prevent misclassification. Therefore, this paper proposes an efficient regularization approach within the Mask Region-based Convolutional Neural Network (Mask_RCNN) optimizer by integrating Weighted Mean League 2 (WMean_L2) with Stochastic Gradient Distance (SGD). We introduce this model as Mask_RCNN+SGD+WMean_L2. WMean_L2 is formulated to ensure consistency in penalty regardless of model size, providing stability across architectures and simplifying hyperparameter tuning. This approach enhances the preservation of discriminant features while achieving consistent and optimal classification performance. We tested our model on the benchmark database known as Beijing Institute of Technology (BIT), evaluating its performance based on precision, recall, F-score, and accuracy. Our results demonstrate significant efficiency improvements compared to previous studies, with precision ranging from 92.31% to 100%, recall from 94.74% to 100%, and F-score from 93.51% to 100% across six vehicle classes, achieving the highest average accuracy of 97.22%.