目前政府對於寵物和流浪動物管理主要採用的方法是植入RFID晶片，用以識別寵物身分。由於採用侵入式手段，民眾為寵物植入晶片意願不高，導致管理漏洞。本研究針對以非侵入的影像辨識方法進行寵物身分識別，提出一個多模態生物辨識方法。我們先藉由CNN分類出犬種，再擷取寵物鼻紋、身形輪廓及臉部幾何等多模態生物特徵，最後藉由混合神經網路分類器進行犬隻身分識別。最後我們設計了一個多模態混合神經網路（MM-HNN）分類器系統來驗證寵物辨識的性能。實驗比較顯示，單純使用CNN深度學習模型從事寵物身分識別的等錯誤率為23.77%，MM-HNN寵物身分識別的等錯誤率為13.45%，而CNN犬種辨識加上MM-HNN身分識別的等錯誤率為4.65%；在三群同胞犬的實驗中，身分識別的等錯誤率為4.57%；在三種生物特徵模態的辨識實驗中，在紋理模態中辨識率為88.33%，輪廓模態中辨識率為84.83%，臉部模態中辨識率為79.83%，最後混合三種模態而成的多模態辨識率則達到95%，因此證明了多模態混合神經網路分類器具有良好的辨識性能。;At present, the government mainly uses embedded RFID chip to supervise pets and stray animals to identify the pets. As the means is invasive, common people′ willingness to embed the chip in pets is not high, leading to administrative vulnerability. This study proposes a multi modal biological recognition method for using non-invasive image recognition method to identify pets. The dog species is classified by CNN (Convolutional Neural Network), and then the pet′s multi modal biological features are extracted, such as muzzle pattern, body contour and facial geometry. Finally, the dog is identified by hybrid neural network classifier. A Multi Modal Hybrid Neural Network (MM-HNN) classifier system is designed to validate the performance of pet identification. The empirical comparison shows that the Equal Error Rate (EER) of simply using CNN deep learning model for pet identification is 23.77%, the EER of MM-HNN pet identification is 13.45%, and the EER of CNN dog species recognition + MM-HNN identification is 4.65%. In the experiment on three groups of fellow dog, the EER of identification is 4.57%. In the recognition experiment on three biological feature modals, the texture modal recognition rate is 88.33%, the contour modal recognition rate is 84.83%, the face mode recognition rate is 79.83%, and the multi modal recognition rate of the three modals is 95%, proving that the MM-HNN classifier has good recognition performance.
