轻量级深度特征交互融合的车辆重识别网络研究OA

Vehicle re-recognition requires the model to focus on both the vehicle's overall outline and the vehicle's sub-tle local details at different stages to extract distinguishing features at a deeper level.To address the above problem,a pyra-mid branch with a lightweight large sense field is constructed,which significantly improves the performance of the back-bone network while introducing only less than 0.84 million additional parameters,which allows the model to focus on the global texture at the deeper level of the network.A backbone guided fusion module is proposed to enable the pyramid branches to learn effective feature representations,which can adaptively fuse the pyramid branch features with the back-bone features to help the pyramid branches learn effective information.In addition,an image deblurring technique is employed to preprocess the input features and combined with a parallel attention mechanism to enhance the attention to feature details.Experiments conducted on the Veri-776 and VehicleID datasets show that the proposed lightweight app-roach effectively improves vehicle re-recognition accuracy and generalization ability.