A novel Jinnan individual cattle recognition approach based on mutual attention learning scheme

Abstract

The accurate recognition of individual Jinnan cattle is crucial for efficient breeding. However, existing ear tagging and machine learning approaches are inefficient, and deep learning models with a significant number of parameters can be computationally expensive. To address these challenges, this study proposes an innovative mutual-attention learning method for efficiently recognizing individual Jinnan cattle. This method employs two attention-enhanced student networks that collaborate and exchange knowledge on cross-dimensional attention during training, achieving improved recognition performance compared to counterparts without mutual learning. Additionally, smaller student networks can achieve improved recognition performance with much less computational cost, leveraging knowledge from larger networks. Furthermore, the proposed approach integrates attention modules such as block attention module (BAM), convolutional BAM (CBAM), and triplet attention units to capture spatial and channel attention for obtaining robust feature representations, enhancing performance. Our proposed model achieves state-of-the-art classification accuracy of 98.91%, surpassing other models in both accuracy and efficiency. Experiments also demonstrate its superiority by improving accuracy by over 2.97% and reducing loss value by over 50% compared to existing methods. Overall, the proposed approach has great potential to revolutionize the recognition of individual Jinnan cattle.