基于多尺度特征融合增强的InceptionResNetV1模型的牦牛面部识别OA

In order to solve the problems of low efficiency in manual recording,easy loss of physical labels,and difficulty in insurance claim assessment in traditional yak identity management,this study proposed a yak face recognition method named YakFaceInception.This model introduced a feature pyramid network(FPN)of improved smooth module at the neck of the InceptionResNetV1 model to facilitate multi-scale feature fusion.The input yak facial image was mapped into 512-dimensional feature vectors through a model,and then the cosine similarity between the feature vectors was calculated to determine whether they belong to the same yak.The results showed that the YakFaceInception model achieved a recall rate of 92.9%,which was slightly lower than that of the ArcFaceNet model.However,its precision,F1 score,and model size were all superior to those of ArcFaceNet,ResidualAttentionNet,Cbam_ResNet_50,MobileFaceNet,and ResNet50.The YakFaceInception model had a recall rate of 92.9%,an accuracy rate of 96.8%,an F1 score of 94.8,and a model size of 115.5 Mb on the test set of the self-built original yak facial dataset.Compared with the original InceptionResNetV1,the recall rate,accuracy rate,F1 score,and model size were increased by 0.54%,1.57%,1.07%,and 391.49%,respectively.The results indicated that comprehensive performance of YakFaceInception model was better than that of InceptionResNetV1 model,and the prediction of the model tends to be conservative.