面向复杂光照条件的煤矿视频增强与目标检测算法研究

打开文本图片集

Abstract：Complex lightingenvironments（suchaslowilumination，strong shadows，anddustinterference）incoal mines seriouslyrestrictthetargetdetectionaccuracyofvideosurveillncesystems.Thispaperproposesannd-toendjointoptiization algorithmthatintegratesadaptivevideoenhancementandanimprovedYOLOv7-tiy.Fistly，amulti-scaleRetinexdecomposition modelisconstructedTheillminationcomponentispredictedbyUNet，andlcalbrightnessequalizationisachievedthrough dynamicGammacorrection.Non-localmeanfiteringiscombinedtosuppessdustnoise.Secondlythecros-stagefeaturefusion （CSFF）moduleandthelight-awareatention （AA）mechanismareintroducedintoYOLOv7-tinytoenhancethesemanticfeature expressioninlow-illuminationareas.Theexperimentsarebasedontheself-builtundergroundcoalminedatasetCMVID-2000. Theresultsshowthattheaverageprecision（mAP）ofthetargetdetectionofthealgorithminthispaperreaches89.7%，whichis （20号 12.3% higher than thatofthe baselinemodel.ThePSNRandSSIMof theenhancedimageare28.6dBandO.91respectively， andthe inference speedreaches 41 FPS，meting thereal-timemonitoring requirementsof underground mines.

Keywords：inteligent mine;video enhancement;target detection;complex lighting learning;end-to-end learning

煤炭作为我国主体能源之一，其安全生产始终是国家战略的重要关切。（剩余4921字）