Wildlife　recognition　is　the　task　of　matching　and　outputting　similar　probabilities　given　two　images,　with　twin　networks　being　a　popular　framework　in　the　field.　In　this　work,　we　use　the　Siamese　architecture　as　the　main　framework.　As　for　the　backbone,　we　compare　two　popular　backbone　networks,　including　VGGNet　and　ResNet,　respectively.　We　compare　those　two　networks　with　respect　to　the　convergence　speed,　robustness　and　maximum　accuracy　to　exploit　the　effectiveness　of　our　method.　We　test　those　networks　on　AFHQ　dataset.　Our　method　with　ResNet　achieves　98.86%　accuracy.　Our　experiments　have　confirmed　that,　to　a　certain　extent,　the　larger　the　number　of　layers　in　the　network,　the　higher　the　stability　and　accuracy　of　model　training.　©　2022　SPIE.