Research On The Construction Of Primate Face Datasets And Recognition Methods

In recent years,datasets that apply to different research tasks have been emerging continuously,the scale is increasing,and the scene is becoming more and more complex.The establishment of the datasets provides a data basis for the development of scientific research.The increasing scale of datasets,and the diversification of the data type,which is also puts forward new challenges to existing theories and methods,thus makes the new algorithm has been put forward with higher performance.However,there is no authoritative datasets for wildlife research in the world at present,and the published individual animal datasets generally have problems of small amount of data and poor picture quality.This greatly restricts the improvement and application of the animal face recognition algorithm.At the same time,in the field of monkey face recognition,current research progress of just stop at the level of reuse of existing face recognition technology.However,the facial structure,organ shape and facial texture of primate wild animals differ greatly from human beings.Therefore,the application of the same model in the face of monkey face is greatly reduced,however,there is no specific network model for facial recognition of primates.According to the above problem,this article first built the first large primate face datasets.In the process of datasets building,a fast and effective primate facial detection method is proposed to quickly obtain the monkey face region.Then,we propose a new CNN model,Fine-Grained CNN(FCNN),for facial recognition of primates.The main work of this paper is as follows:1.A large primate face dataset constructed for the lack of facial database in individual primate identification.We have explored the corresponding data acquisition methods and strategies,and finally completed the establishment of datasets.The dataset included 43 primate species,1040 individuals and 102,399 clear facial images.The average number of images per individual is about 100.The species collected include golden monkeys,gorillas,baboons,etc.Finally,the existing methods are implemented to verify that the database can be applied to the individual identification,gender identification and age recognition of primates.2.An automatic detection method based on CNN is proposed in order to quickly establish a primate face datasets.In this paper,the entire algorithm flow is presented,quantified by HSV color space of the suspected area of primate faces in the original image,and then use facial data training CNN model,to complete the monkey face selection and positioning,finally complete the monkey face detection.3.A modified CNN model,FCNN,is proposed to identify the fine particle size of the primate face.There is a lack of both low training cost and high accuracy methods for face recognition in primates.In view of the high similarity of the primate face's overall structure and the fact that it is covered with hair,it is necessary to take into account the feature extraction of hair and facial features simultaneously.The FCNN uses 3×3 convolution kernels and three 3×3 convolution kernels interchangeably instead of traditional CNN convolution kernels.The experimental results show that this method can be used for face recognition of primates in various scenes quickly and effectively.