全球各地比特犬襲擊人畜新聞頻繁出現，各國政府開始訂定比特犬禁養或列管政策，比特犬自動辨識就成了行政上必要的工具。本研究提出了一種基於體型/狗臉雙模態特徵辨識方法來識別比特犬種，我們使用Segment Anything Model(SAM)網路模型來獲取比特犬體型外觀和狗臉範圍，再利用ResNet進行遷移式學習，同時學習比特犬種的臉部外觀特徵與體型外觀特徵並進行融合決策辨識。在實驗方面，我們根據不同應用場景設計三個實驗來驗證雙模態辨識方法的可靠性，分別是比特犬種與非比特犬種的辨識實驗、管制犬與非管制相似混血比特犬的辨識實驗以及管制犬與非管制犬的辨識實驗，實驗結果顯示在比特犬種與非比特犬種之間的鑑別，雙模態辨識方式則可達到97.4%的辨識準確率；而在管制比特犬與非管制之比特混血犬種之間的辨識，雙模態辨識方式能夠達到90.2%辨識準確率；最後在管制犬與非管制犬之間的辨識，我們的方法能夠達到87.84%的辨識準確率。;Frequent reports of human and livestock attacked by pit bulls have led governments worldwide to establish policies prohibiting or regulating the ownership of pit bulls. Automatic recognition of pit bulls has become a necessary administrative tool. In this paper, we propose a dual-modal feature recognition method based on body size and dog face to identify pit bull breeds. We use the Segment Anything Model (SAM) network model to extract the appearance and facial regions of pit bulls, and then employ ResNet for transfer learning. The method simultaneously learns facial and body appearance features of pit bull breeds and performs fused decision recognition. In terms of experiments, we design three experiments to verify the reliability of the dual-modal recognition method in different application scenarios. These experiments include the recognition of pit bull breeds versus non-pit bull breeds, the recognition of regulated pit bulls versus non-regulated pit bull mixes, and the recognition of regulated dogs versus non-regulated dogs. The experimental results demonstrate that the bimodal recognition approach achieves a recognition accuracy of 97.4% in discriminating between pit bull breeds and non-pit bull breeds. For the recognition of regulated pit bulls and non-regulated pit bull mixes, the bimodal recognition approach achieves a recognition accuracy of 90.2%. Lastly, in the recognition of regulated dogs and non-regulated dogs, our method achieves a recognition accuracy of 87.84%.
