Welcome to Smart Agriculture 中文

Smart Agriculture ›› 2022, Vol. 4 ›› Issue (2): 77-85.doi: 10.12133/j.smartag.SA202201001

• Topic--Smart Animal Husbandry Key Technologies and Equipment • Previous Articles     Next Articles

Multiscale Feature Fusion Yak Face Recognition Algorithm Based on Transfer Learning

CHEN Zhanqi1(), ZHANG Yu'an1(), WANG Wenzhi1, LI Dan1, HE Jie1, SONG Rende2   

  1. 1.Department of Computer Technology and Application, Qinghai University, Xining 810016, China
    2.Animal Disease Prevention and Control Center of Yushu Tibetan Autonomous Prefecture, Yushu 815000, China
  • Received:2022-01-02 Online:2022-06-30 Published:2022-08-04
  • corresponding author: ZHANG Yu'an E-mail:1046788801@qq.com;2011990029@qhu.edu.cn


Identifying of yak is indispensable for individual documentation, behavior monitoring, precise feeding, disease prevention and control, food traceability, and individualized breeding. Aiming at the application requirements of animal individual identification technology in intelligent informatization animal breeding platforms, a yak face recognition algorithm based on transfer learning and multiscale feature fusion, i.e., transfer learning-multiscale feature fusion-VGG(T-M-VGG) was proposed. The sample data set of yak facial images was produced by a camera named GoPro HERO8 BLACK. Then, a part of dataset was increased by the data enhancement ways that involved rotating, adjusting the brightness and adding noise to improve the robustness and accuracy of model. T-M-VGG, a kind of convolutional neural network based on pre-trained visual geometry group network and transfer learning was input with normalized dataset samples. The feature map of Block3, Block4 and Block5 were considered as F3, F4 and F5, respectively. What's more, F3 and F5 were taken by the structure that composed of three parallel dilated convolutions, the dilation rate were one, two and three, respectively, to dilate the receptive filed which was the map size of feature map. Further, the multiscale feature maps were fused by the improved feature pyramid which was in the shape of stacked hourglass structure. Finally, the fully connected layer was replaced by the global average pooling to classify and reduce a large number of parameters. To verify the effectiveness of the proposed model, a comparative experiment was conducted. The experimental results showed that recognition accuracy rate in 38,800 data sets of 194 yaks reached 96.01%, but the storage size was 70.75 MB. Twelve images representing different yak categories from dataset were chosen randomly for occlusion test. The origin images were masked with different shape of occlusions. The accuracy of identifying yak individuals was 83.33% in the occlusion test, which showed that the model had mainly learned facial features. The proposed algorithm could provide a reference for research of yak face recognition and would be the foundation for the establishment of smart management platform.

Key words: yak, face recognition, transfer learning, feature pyramid structure, T-M-VGG

CLC Number: