Research On Active Millimeter Wave Sequence Image Target Detection Method Based On Deep Learning

Due to the frequent occurrence of terrorists carrying dangerous goods to endanger public security,human security is now highly regarded in the world.The active millimeter wave image detection method can detect the human body better,but there are still many problems such as the privacy of the inspected person and the visual fatigue of the inspector in pedestrian detection.The target detection method based on deep learning can realize the automatic learning of target characteristics and perform end-to-end human detection.Therefore,this paper will study the corresponding hidden object detection algorithm based on the characteristics of the active millimeter wave image data set used.The main research contents are as follows:(1)The existing problems in detecting hidden objects in millimeter wave images using traditional image processing methods and target detection methods based on deep learning are analyzed.It points out that the active millimeter-wave image itself has a low signal-to-noise ratio,little information and artifacts,which leads to insufficient detection performance.It also mentions the limitations of detecting single frame images.(2)In view of the above problems,combined with the powerful feature extraction ability of convolutional neural network and the processing ability of cyclic neural network to deal with temporal features,a detection network based on temporal multiple angles is constructed.At the same time,the experimental results are compared by combining different network structures,and their advantages and disadvantages are analyzed to lay a foundation for the improvement in the following paper.(3)In order to reduce the false alarm rate of the detection network and improve the detection performance,mixup data enhancement is used in the data pre-processing part to amplify the data set and improve the model generalization ability.At the same time,it combines the attention mechanism and the online hard example mining technology to improve the sensitivity of the model to difficult-to-resolve samples and weaken the influence of artifact noise on the detection results.Finally,it is found that the detection performance is indeed improved through the experimental comparison,and the best detection model can be selected to pave the way for the subsequent integration learning.(4)Different detection models are constructed through different input angles,and integrated learning method is adopted to integrate such models,so as to reduce the error judgment caused by the detection errors of a single model,enhance the stability of detection,and improve the overall detection performance.This article mainly uses the Xgboost algorithm for integration,and conducts a comparative test with the voting integration method to analyze its advantages and disadvantages.