面向机器视觉任务的多尺度图像特征压缩算法

Fortheproblemsthat inthecollaborativeintelligenceframework theintermediate featuredataof machinevisiontasksislargeanddifficulttotransmitefficiently,a multi-scale imagefeaturefusioncompressionalgorithm wasproposed.Firstly acascadedresidualtransformationmodule wasdesignedaccordingtothe multi-scalefeaturesoutputbythedeeplearning modelontheedgedevice theredundancyofmulti-scalefeatureswaseliminatedbystepwisesubtractionoffeaturesofdifferentsizes andtheresidualfeatureswerecompressedtoaunifiedsize. Then anautoencoderwasdesignedtoeliminatethestatisticalredundancyofcompactfeaturesby arithmeticcoding.Next a prediction andreconstruction module was designed onthecloud accordingtothecompactfeaturesofdecodingtogeneratethepredictionfeatures which were combinedwiththeresidualfeaturestoaccuratelyreconstructthemulti-scalefeatures.Finally,a jointoptimizationfunction wasbuiltforthecollaborativeoptimizationofthe modulesincluding residualtransformation autoencoder andpredictionreconstruction thusachievingtheoptimal trade-offbetweentransmissionbitrateandinformationrepresentationability.Thesimulation resultsshowthattheproposedalgorithmhasnotonlythelargestfeaturecompressionratio but alsothemostcompletereconstructedfeaturesinthespacecompression andthatwhenthetransmissionbitrateis0.1bpp themodelaccuracyoftheproposedalgorithmisimprovedby8.57% and3.87% respectively comparedwiththeimagecodingalgorithm VVCandthefeaturecompressionalgorithm MSFC.Thisstudycanprovidetechnicalsupportforthecodingframeworkof machinevision andhascertainvalueinengineeringapplication.