| Summary: | Individual cattle identification is pivotal for dairy farming, food quality tracing, disease prevention and control, and registration against fraudulent insurance claims. When employing neural network models for cattle face recognition, challenges arise due to limited individual data, varying cattle face positions and angles, and significant image background noise. This often results in the model’s low robustness in recognizing untrained individuals. To address this, we introduce an algorithm based on the Siamese Group Chunking (GC) Capsule Network (SGCCN). Firstly, the GC block serves as the feature extractor for the primary capsules. By utilizing separate filters, the GC block learns the unique representations of cattle faces, enhancing feature extraction capabilities while reducing model parameters. Secondly, a adjusted cosine similarity is employed to capture both directional and absolute numerical differences between cattle face vectors, bolstering the network’s robustness. Experimental results reveal that, compared to the conventional Siamese capsule network, the SGCCN reduces parameter usage by 57.65% yet increases recognition accuracy by 7.67%. The recognition rate on the validation set reaches 92.67%, and 89.33% for untrained individuals.
|
|---|