Structural description model for vehicle feature recognition

Precise vehicle recognition has long been neglected compared with face recognition, meanwhile, benefiting from the development of neural networks, as same as face recognition, structural vehicle feature recognition is currently feasible. In this paper, we use a CNN-based cascaded multi-task framework for vehicle detection and alignment, then we trained a backbone CNN which can learn a mapping from vehicle image to a Euclidean space. Therefore, task of vehicle recognition can be solved as face recognition. Besides, different vehicles have different attributes compared with faces, we enriched the largest open source vehicle recognition dataset VehicleID with color and direction while the branch-CNN is employed to learn multiple features from different branches, outputs. Finally, structural vehicle features can be transformed from image to text which enhances the data expression ability.