Research On Target Detection Method Based On Improved FPN Algorithm

Target detection is an important computer vision task for classifying and locating objects in an image.In recent years,with the rapid development of deep learning technology,target detection has ushered in a new development opportunity,and many classical algorithms have been produced.Generally speaking,the target detection algorithm can be divided into two branches from the perspective of whether candidate frames are generated or not: the single-stage detection algorithm with good real-time performance and the algorithm based on region proposal with relatively excellent accuracy.Multi-scale target detection is always the basic challenge of these two algorithms,especially for small target detection.In this paper,the target detection algorithm of feature pyramid networks(FPN)is studied,and two optimization algorithms based on FPN are designed.One is to simulate the receptive field mechanism of the organisms by introducing the RFB(Receptive Field Block)module,which makes the network focus on learning the features located in the center,and then obtain better detection results.Another improvement method is to introduce the prediction optimization module for the first time.The module uses the context information from the region of interest to make the features have stronger semantics,then uses the internal cascade network to implement multi-stage detection,and the threshold value of the cascade network will gradually increase,so as to continuously improve the positioning ability of the algorithm.In addition,the two improved methods are trained by the network model Det Net-59.The network is specially designed for target detection.On the premise of ensuring the depth of the model,the network further increases the resolution of the feature map,and then improves the detection performance of the algorithm.After training on standard data sets VOC,the two improved algorithms proposed in this paper achieve 80.2% and 80.9% accuracy respectively without any optimization method,which shows that the algorithm has good detection performance.Finally,by using the classical face algorithm and the improved algorithm proposed in this paper,and building the server and database,the intelligent monitoring system in classroom scene is realized.The system achieves good detection results for students' behavior and face detection,which shows that the improved algorithm proposed in this paper has certain practical value.