Research On Vehicle Component Detection And Re-identification Based On Improved EfficientDet Algorithm

Vehicle component detection and re-identification using deep learning has become an interesting research field as computer vision technology and deep learning advance.This area plays a crucial role in theoretical research in fields such as traffic management,intelligent parking,security monitoring,and autonomous driving.It holds broad prospects and practical value in application fields.However,vehicle component detection and reidentification are challenging tasks due to the scale transformation of vehicle components,viewing angle,and illumination.Hence,effective extraction of robust and recognizable detail features in images is key to improving the detection and re-identification of vehicle parts.This thesis proposes an efficient and accurate vehicle component detection network,EFDet-SPP,based on the deep learning network Efficient Det.Firstly,a Bi FPN network with vertical cross-layer connection is designed for feature fusion,inspired by the idea of cross-layer connection and residual network.This comprehensively balances the input streams of high-level nodes and low-level nodes,shortens the distance from the low-level feature map to the top-level feature map,and realizes better information interaction.At the same time,anchor box prediction is transformed into pixel prediction to overcome interference caused by changes in the model application scene,eliminating hyperparameters related to the anchor box and reducing the calculation amount,which improves applicability in the vehicle component detection scene to a certain extent.To tackle the challenge of tiny vehicle components,the combination of mosaic and copypaste data enhancement methods balances the samples and improves the network’s generalization ability.High semantic information is captured effectively by using spatial pyramid pooling in the feature extraction network.Due to the lack of public datasets,we have established two vehicle component detection datasets VLC and VDC to ensure sufficient data and rich scenes.Moreover,this thesis focuses on finding local features with discrimination without ignoring the global information of the whole vehicle image.A multi-granularity vehicle re-identification algorithm based on component and global features is proposed.Local features are extracted from component images captured by the simplified EFDet-SPP,and global information is extracted from the whole image by performing multi-scale feature extraction.By doing this,global features of the vehicle and local features of the components with high discrimination are both effectively learned.More importantly,we design a novel feature learning strategy for more fine-grained features,where the discriminant information is divided into granular features of different sizes.At the same time,instead of learning in the semantic region,the image features are evenly divided into several stripes from the vertical direction,and the number of stripes in different local branches is changed to obtain a multi-granularity local feature representation.To test the proposed method,we conducted comparative experiments on datasets VLC and VDC,proving that EFDet-SPP can achieve efficient and accurate vehicle part detection.We used the public datasets Ve Ri-776 and Vehicle ID to conduct vehicle reidentification experiments.The final experimental results verified the effectiveness of the proposed method for vehicle component detection and re-identification tasks.