Improving Vehicle Identification Through Advanced Fine-Grained Vehicle Classification

Vehicle identification plays a crucial role in Intelligent Transportation Systems, impacting areas such as toll collection, vehicle access control, and criminal forensics. Despite recent strides in Automatic License Plate Recognition (ALPR) research, real-world scenarios still pose significant challenges. This work explores potential enhancements in vehicle identification systems by integrating modules such as ALPR with Fine-Grained Vehicle Classification (FGVC), which categorizes vehicles based on attributes such as type, make, model, and year. Our study focuses on advancing FGVC, particularly vehicle type classification. We investigate selective prediction, a technique that allows models to discard uncertain predictions, and examine superclass methods, including a novel online superclass approach that operates solely during the test phase. We trained and evaluated four deep learning models using a dataset adapted from a widely adopted ALPR dataset. The results demonstrate that both superclass methods and selective prediction improve classification accuracy, with the combination of online superclass and selective prediction delivering the best performance. Future research will focus on integrating these enhancements into ALPR systems to determine how FGVC can further enhance their capabilities.