A Feature Fusion based Custom Deep Learning Model for Vehicle Make and Model Recognition

Vehicle Make and Model Recognition (VMMR) is a pivotal task in various domains including surveillance, traffic management, and the automotive industry. Despite significant progress in deep learning approaches, existing VMMR systems still struggle to attain robust accuracy, particularly across diverse environments and viewing angles. In this paper, we propose a novel approach based on the fusion of two feature maps extracted from DenseNet201 and ResNet50V2 baseline models, respectively. Additionally, we employ a modified Convolutional Block Attention Module (CBAM) to enhance the resilience and precision of VMMR systems. By incorporating attention mechanisms in the feature extraction process, our modified CBAM model effectively captures both spatial and channel-wise dependencies, facilitating more potent discriminative feature representations. We assess our proposed approach through extensive experiments on two benchmark datasets namely Stanford Cars and CompCarsSV. Achieved accuracies are 93.51% and 99.03%, on Stanford Cars and CompCarsSV, respectively that are better than past methods. The code of our proposed model can be found at: https://github.com/JUVCSE/featurefusion.