A lightweight deep learning model for cattle face recognition

At present, there have been some researches on deep neural networks used in the field of biometrics to solve the problem of cattle identity authentication, particularly facial recognition under non-invasive methods. However, due to the large scale of deep neural network models, it is difficult to implement directly in resource-constrained embedded systems. Therefore, a lightweight neural network is designed in this paper that can be deployed in embedded systems that requires a small amount of weight representations and low-cost operators. The lightweight neural network designed in this paper contains six convolutional layers, which the convolution with a step size of two is used to achieve pooling, the Batch Normalization algorithm is used to normalize the neural network input, the dropout layer is used after the activation function Relu, and global average pooling is used to replace the fully connected layer after the sixth convolutional layer. The final parameter is 0.17 M, the model size is 2.01 MB, and the MFLOPs is 9.17. We have collected 10,239 cattle face images (103 subjects) from actual cattle farms to construct a dataset. The experimental results show that the accuracy of our proposed model is 98.37%, the macro average is 98%, and the Grad-CAM algorithm is used to verify the adequacy of features extracted by this network and avoid extracting noise features. The model we proposed is transplanted to the Raspberry Pi after training on the PC to test the running time, the model we proposed achieves a minimum detection time of 300 ms per picture. Our experiment results show that the lightweight neural network we proposed can achieve high recognition accuracy for cattle face recognition and significantly reduce the computational cost.

Automated summary

A lightweight neural network is designed in this paper for high accuracy cattle face recognition in embedded systems, significantly reducing computational cost.