Research On Vehicle Multi-Attribute Recognition Method Based On Deep Learning

In recent years,the increase in vehicles has brought more convenience and caused more traffic problems.The realization of functions such as vehicle diversion and vehicle tracking in intelligent transportation systems requires the detailed attributes of the vehicles as information support.The traditional target recognition detection model is huge,low in real-time,and has high requirements on the hardware environment.The attribute recognition of the vehicle is only a single attribute such as color,car model,license plate,etc.,which cannot meet the existing traffic system.In order to improve the effect of vehicle detection and attribute recognition in actual scenes,a multi-attribute vehicle recognition method is proposed based on deep neural network,which detects vehicle targets and recognizes vehicle multi-attributes in modules.In the target detection module,in order to overcome the impact of the particularity of vehicle images on feature extraction and ensure the efficient operation of the network on mobile devices,based on the YOLOv4 algorithm,a lightweight target detection algorithm(GhostNet Lite-You Look Only Once version 4)is proposed.The backbone network is replaced with the GhostNet structure,and the amount of calculation and parameters are reduced through the GhostNet residual structure.In order to improve the detection accuracy,the original PANet structure is replaced by the Bi-PANet structure that integrates the Bi-FPN idea,and the feature fusion is improved through cross-layer weighted bidirectional fusion.Finally,the convolution structure is replaced,and the deep separable convolution is used to deepen the lightweight level of the network.In the fine-grained classification module,vehicle targets have diversified scales,random occlusion incompleteness,and uncertainties such as illumination and weather.The improved BCNN algorithm is used to achieve fine-grained feature extraction,and vehicle attributes are classified according to vehicle color and vehicle.It is divided into three aspects: orientation and vehicle type.First,select a good feature extraction sub-network.By comparing multiple subnetwork structures,the two models with high total detection accuracy are selected as the most algorithmic sub-network structure to ensure the efficient feature extraction of the network.Subsequently,by replacing the activation function in the sub-network,the h-swish function is introduced to accelerate the convergence speed of the model and improve the detection accuracy of the model.Finally,the two modules are operated jointly to observe and verify that the system has good detection performance.In the target detection module,test experiments on the joint data set of VOC2007 and VOC2012,the average accuracy rate(m AP)of the G-YOLOv4 algorithm reached 80.82%,the model size was reduced to 43.1MB,and the vehicle accuracy rate(AP)after migration learning reached It achieved 92.99%,and the FPS reached 42.89.While achieving high precision,it also improved the real-time performance of detection and completed the lightweight improvement of the network.In the fine-grained module,the Vehicle ID data set is used for test experiments,and the vehicle type is divided into 17 types to realize the recognition of multiple attributes of the vehicle.The improved BCNN algorithm has a total detection rate of 93.692%,and it has a good detection effect in scenes such as day,night,rain,and tilt.Finally,combined with the output results of the two modules,the effectiveness of the multi-attribute recognition detection system is verified.