生猪背票厚度无接触检测方法研究

打开文本图片集

中图分类号：TP391.4 文献标识码：A 文章编号：2095-5553（2026）02-0283-10

Abstract：Existing single-modality methods fordetecting pig backfatthicknessoftenoverlookspig body measurement data，resulting infailure tocapturetherelationshipbetweenthethree-dimensionalcharacteristicsof thepig's back andits backfatthickness.Consequently，thesemethods demonstrated limitedgeneralizabilityandlowerdetectionacuracy.To address this issue，thisstudy proposedanon-contact detectionapproach basedonmultimodal data fusion.Pig body measurements were extracted through image registration and coordinate transformation techniques.Three data modalities： bodymeasurementdata，depth imagesof the pig's back，andRGB imageswere integrated toconstruct sevendistinct datasets.The detection performanceof models utilizing single，dual，and multiple modalities was compared.Toenhance the model’sreceptivefieldandfeatureextractioncapabilities，theLarge Selective Kermel（LSK）andOmni-dimensional Dynamic Convolution（ODConv）were introduced.Additionally，a Self-attention Deep Equilibrium MultimodalityFusion （SDE）algorithm was proposed toaddress thechallenges of cross-modality interactionand feature lossoftenencountered inconventional fusionstrategies.Experimentalresultsshowedthat incorporatingaditionaldatamodalitiessignificantly improveddetectionaccuracy.With theintegrationofLSK，ODConv，and SDE，the model’smeanabsoluteerror

0 .MAE， ），root mean square error（RMSE），and mean absolute percentage error （MAPE ）werereduced by 30.94% 28.73% ，and 29.82% ，respectively，reaching 0.36mm ， 0.57mm ，and 2.27% ．Moreover，the coefficient of determination （ R2， ）increased by 6.07% ，achieving a value of O.94. This multimodal fusion approach not only meets the accuracyrequirementsforpracticalpig backfatthickness detectionbutalsolaysastrong foundationforfurthertechnological advancements in precision livestock farming.

Keywords：pigs；backfat thickness；multimodality fusion；deep learning；feature extraction

0 引言

中国是世界上最大的生猪养殖国，猪肉产量占肉类产量近 2/3[1] ，监测生猪体况可大幅提升生产效率。（剩余15203字）