Research On The Detection And Recognition Algorithm Of Key Information Of The Network Contract Vehicle Based On RetinaNet

With the popularity of online car travel in people’s daily life,the safety of online car travel has become a hot spot.As the cornerstone of solving the safety problem of the network contract car travel,the detection and identification of the key information of the network contract car plays a very important role.The current key information detection and recognition algorithm of the network contract vehicle has problems such as low detection accuracy,single algorithm function and slow running speed on the mobile side under the complex application scene.Based on this,this paper studies the key information detection and recognition algorithm based on RetinaNet,detects the vehicle frame containing the head of the vehicle,the vehicle frame with the tail and the license plate,and identifies the door switch status and the direction of the image in the input image.For these target functions,the high-performance algorithm models of server-side and mobile-side are designed to meet the needs of different scenarios.This paper makes an in-depth study on the following problems.First of all,this paper takes RetinaNet as the basic framework of the key information detection and recognition algorithm of the network contract vehicle,and in view of the problem of low detection accuracy in the current practical application scenario,this paper improves the RetinaNet algorithm from the aspects of network structure,loss function,and candidate box(Anchor)selection method.The improved server-side algorithm detection accuracy was improved by 1.1 percentage points compared with the original RetinaNet algorithm,and the detection accuracy reached99.9%.Secondly,this paper embeds the branch that recognizes the state of the door switch and the branch that recognizes the direction of the shooting image on the framework of the key information detection and recognition algorithm of the net-contract vehicle,and realizes the single-model multi-functional high-performance network key information detection and recognition algorithm.The test results show that the key information detection and recognition algorithm designed in this paper can reach 99.8% of the newly embedded identification branch accuracy under the condition that the key information detection accuracy of the original network contract vehicle is 99.8%.Finally,in view of the problem that the mobile end compute force level is limited and the storage is small,the model performance is reduced,this paper proposes to take Mobile Net V3 Small as the basic feature extraction network of the key information detection and recognition algorithm of the mobile end network contract vehicle.Channel compression and embedding attention mechanism for feature multiplexing modules to improve the performance of mobile models.On the surface of the test results,the mobile-end algorithm model proposed in this paper can reach100 ms in the inferred speed on mobile devices,which meets the requirements of real-time scenes,with the comprehensive accuracy rate of the key information detection and recognition of the network car reaching 98.3%.