基于自适应融合CNN一OF特征和LSTM网络的猪攻击行为识别

打开文本图片集

中图分类号：S126；TP391.4 文献标识码：A 文章编号：2095-5553（2026）02-0275-08

Abstract：To recognize aggressive behavior group-housed pigs，analgorithm basedon adaptively fused CNN—OF （convolutional neural network-optical flow）featuresLSTM（Long short-term memory）network wasproposed. 8nursery pigs/pen were mixed in2 pens for3days then8h video was recordedasdatasets in eachday.From the 3-day video pen 1，1 2O0 aggressive1 s episodes 1 200 non-aggressive 1 s episodes were labelled， 80% these episodes were selected as the training set， the remaining 20% asthe validation set. From the 3-day video pen 2， 1 254 aggressve1 sepisodes 85146 non-aggressive1s episodes were labelledas the testset.Firstly，the Horn一 Schunck（HS）method wasused tocalculate the magnitude orientation opticalflow（OF）， then theorientation range optical flows was dividedaccording to the dimension CNN feature maps.Secondly，the histogram the magnitude optical flows was counted in each orientation range， then this histogram wasconverted into a feature map byspatialdimensiontransformation.Finaly，thisfeaturemapwasadaptivelyfusedwiththeCNNfeaturemapbyweight superposition， then the fused features were input into LSTM network to recognize aggression.The accuracy using VGG16—OF—LSTM，ResNet50—OF—LSTM， InceptionV3—OF—LSTM Xception—OF—LSTM algorithms to recognize aggressive behavior pigs was 97.5% ， 97.8% ， 97% 99.3% ，respectively.The result indicates that the CNN—OF—LSTMalgorithmcan beused torecognize aggesive behavior pigs.Furthermore，the proposed adaptive feature fusion approach CNN—OF hasa certain generality.

Keywords：group-housed pigs；aggresionrecognition；convolutional neural network；optical flow；adaptivefusion；long short-term memory

0 引言

在商业猪生产中，不同生长期重新混养是标准操作，这会导致猪攻击直至新的社会等级制度建立[1]。（剩余13185字）