Object Perception And Application Based On Ensemble Learning

The performance of object perception is particularly important with the advent of the digital information torrent era.However,traditional models are difficult to meet the performance of object perception when facing perceived objects in different manifestations.Therefore,the perception scheme based on ensemble learning is more and more important due to its superior performance.This paper focuses on three scenes of object perception,digital image device perception,visible image content perception,and SAR image signal perception.We study the applicability of object-aware application and ensemble learning algorithms for specific perceptual performance in depth.A perception performance improvement scheme based on ensemble learning is proposed for the objects of three manifestations.The main research work of our paper is as follows:(1)For digital image device perception,we proposed the real-time camera source identification methods based on ensemble of multi camera fingerprints.We establish the database of 100 video clips originated from 25 mobile phones.Based on this,we proposed a real-time device-aware computing method for video stream and final ensemble solution.The ensemble solution significantly improves the performance of the accuracy of camera source identification and achieves accuracy of 98.161%for Android phones,which significantly higher than single models.(2)For visible image content perception,we proposed deep ensemble scheme based on active and passive diversity constraints.We analyzed the problem of parameter instability for active deep ensemble constraints.And the diversity of network is generated based on passive deep ensemble constraints.A distillation scheme is proposed for deep ensemble with passive diversity constraints in order to resolve the problem of large amounts of parameters in deep ensemble networks.Futhermore,in order to resolve the problem of too many training times in the traditional deep ensemble network distillation scheme,we innovatively proposed the Lazy Born Again Network.This proposed Lazy Born Again network is based on the learning model of cyclic learning rate,and it used the historical network snapshot generated during the training process to distill the following training process.The proposed Lazy Born Again network achieves comparable performance with traditional integrated distillation networks in a single training process,(3)For SAR image signal perception,we proposed the Flexible py-StackNet(FPS)ensemble learning framework.The framework divides the original Stacking ideas into learning layer,data stream layer and evaluation layer.Based on learning layer,it supports multiple models with complete learning solution.Based on data stream layer,it supports flexible feature combination and feature flow.Based on evaluation layer,it supports real-time evaluation and adaptive model selection.The SAR image signal perception scheme based on the FPS framework won the championship of the SAR image classification track of the first space star cup software contest Based on our experiments,effectiveness of the FPS framework is validated.