FREAD: a Feature-aware Region-of-interest Ensemble-based Attention-enableD framework for Class-Incremental Learning

In this thesis, we propose a neural network architecture for fine-grained vehicle make and model recognition from vehicle images. The proposed framework incorporates strategies to improve classification accuracy and integrates incremental learning techniques that enable the model to continuously learn new classes without requiring full retraining. This capability addresses the practical challenge of accommodating newly introduced vehicle models over time. The resulting system has direct applications in intelligent transportation systems, automated highway toll collection, and criminal investigation support, where accurate and scalable vehicle identification is essential.