Dynamic_Bottleneck Module Fusing Dynamic Convolution and Sparse Spatial Attention for Individual Cow Identification.

Main Methods:

• Designed a novel Dynamic Convolution (Dy_Conv) module and integrated it into a Dynamic_Bottleneck module with Sparse-shift Attention (S2Attention).
• Modified the Resnet50 architecture by replacing bottleneck layers with the Dynamic_Bottleneck for enhanced feature extraction.
• Incorporated Query Adaptive Convolution (QAConv) for scale adaptability and a Normalization-based Attention module (NAM) for feature fusion.

Main Results:

• The proposed model achieved high performance metrics on public datasets: Rank-1 accuracy of 96.8%, Rank-5 accuracy of 98.9%, and mAP of 95.3%.
• The model effectively captures and integrates multi-scale features of cow appearance, improving identification accuracy in complex environments.
• Demonstrated superior performance compared to existing methods, particularly in distinguishing visually similar individual cows.

Conclusions:

• The developed model offers a robust and accurate solution for non-invasive individual cow identification.
• The novel modules (Dy_Conv, Dynamic_Bottleneck, QAConv, NAM) significantly enhance the model's ability to handle scale variations and complex scenes.
• This advancement supports precision livestock farming by enabling more effective monitoring and management of individual cows.