基于FNM-Net的輕量級(jí)遙感目標(biāo)檢測(cè)算法

打開文本圖片集

Lightweight remote sensing object detection algorithm based on FNM-Net

WENBin2，ZHANG Jun1，2 ，WANGJunyin2，WANGZihao2，DINGYifu （1.HubeioalbodtfadropisUst; 2.SchoolofElectricaland NewEnergy，ChinaThreeGorgesUniversity，Yichang 443oO2，China; 3.NanchuanPowerSupply Branch of StateGridChongqing ElectricPowerCompany，Chongqing 4O4o，China）

Abstract：Inviewofthe current challnges inremote sensing object detection，such as low accuracy，slow processing speedsandlargequantityof modelparameters，anFNM-Net，alightweightremotesensingobject detectionnetwork basedonan improvedYOLOv7-tinyarchitecture，isproposed.Firstly，alightweightfeatureextractionnetworkFaster-Netisintroducedto substitutetheoriginalone，soastopreventthenetwork'sexcesiveoverlapoffeaturemaps.Secondlythefocalmodulation moduleisintroducedandthespatialinformationintegrationmodule （SIM）isproposed toconstructanewpathaggregationnetwork thatmitigatesissuesoftheinformationredundancyandtheoverlookof intra-layerfeaturesduringfeaturefusion.Andthen，the multi-fine-graineddetectionheadsaredesignedtoavoidthelargescalevariationsofremotesensingobjects.Finalyapruning method utilizinga layeradaptiveamplitudepruning （LAMP）scoreisemployed to eliminateconnections withminor weights， therebyreducingparameternumberandcomputationalburdenandincreasingthedetectionspeed.Thismethodisvalidatedbythe public RSOD dataset.The results show that the proposed method achieves a 51.2% reduction in parameter number，a 55.2% （20 decrease in FLOPs，a 6.5f/s increase in detection speed，and a 2.1% improvement in mAP （mean average precision） in comparisonwiththoseof thebaseline model.Additionaly，itsgeneralizabilityisconfirmedontheNWPU VHR-1Odataset.

Keywords：remote sensing object detection;FNM-Net; light weight; pruning; improvedYOLOv7-tiny; SIM

0 引言

遙感圖像廣泛應(yīng)用于環(huán)境監(jiān)測(cè)、礦產(chǎn)資源勘探、交通規(guī)劃、災(zāi)害監(jiān)測(cè)與救援、國(guó)防軍事等多個(gè)領(lǐng)域。（剩余11290字）