Facial Recognition Of Golden Monkeys Based On Hierarchical Ensemble Network And Self-supervised Clustering

Rhinopithecus roxellana facial recognition is a necessary pre-requisite for golden monkey behavioral recognition research.Individual identification of golden monkeys based on facial traits,on the other hand,has numerous difficulties,including the similarity of individual facial features among golden monkeys and the difficulty of effectively labeling and identifying new individual identities.To that purpose,Hierarchical Ensemble Networks(HE-Nets)and a self-supervised New Individual Recognition(SS-NIR)algorithm based on Siamese networks and deep feature clustering algorithms are proposed in this research.The following are the key research findings:(1)This thesis proposed a monkey face recognition model based on HE-Nets,which is a multi-level ensemble network model composed of multiple subnetworks,inspired by the basic principle of human target recognition,to solve the problem that the facial features of different golden monkeys are very similar,making it difficult for traditional recognition algorithms to accurately recognize them.Each category in the dataset is given a weight by the network model,and each image is given the same weight as its related category.According to the validation data set’s prediction findings,the sub-network will increase the weight of incorrect image recognition and reduce the weight of correct image recognition during the training phase.HE-Nets solve the problem of computing resource waste caused by traditional algorithms that do not account for differences in image recognition difficulty,and the triplet loss function designed in the model also solves the problem of great facial similarity among different golden monkeys,resulting in improved recognition performance.(2)The image data of new individuals cannot be directly identified by the classic individual recognition model for the problem of golden monkey new individual recognition,thus it must be re-labeled and network parameters changed,which is time-consuming and costly.SS-NIR is proposed in this research to overcome the challenge of identifying new golden monkey individuals.Because there is no label information,the network must rely on its own learning ability to identify new individuals,which is inefficient.To lessen the fuzziness of unlabeled data clustering,SS-NIR employs prior knowledge of labeled data sets to learn and identify which attributes of labeled data can form a cluster well.The use of twin networks to train deep neural networks as feature encoders to learn feature distribution of labeled data is discussed in this research.Using prior knowledge from the tagged dataset,the number of new golden monkeys was then approximated.Experiments demonstrate that the difference between the estimated and true values is smaller than three.After adding the anticipated number of individuals,the new individual data of golden monkeys was clustered using res NET-50 and K-means alternate training to complete the identification of the new individual golden monkeys.Individuals with excellent clustering accuracy were finally added to the original golden monkey data set.