- 學位論文
基於Xception深度學習於犬隻品種及身份辨識之研究與應用
Application of Dog Breed and Identification Using Xception-Based Deep Learning Method
摘要
隨著城市快速發展、飼養寵物的人口增加，以及「流浪動物零安樂死」政策之通過，流浪狗問題受到越來越多人的重視。除此之外，寵物走失是飼主們最擔心、著急的事，也因此犬隻辨識成為近期熱門的研究議題之一。 本篇論文提出了一個基於Xception深度學習方法的犬隻品種及身份辨識模型，利用影像辨識的技術尋找走失的寵物犬，本架構會先對犬隻影像做品種之分類，進而於該品種內進行身份辨識。本研究建立了含有十個犬種及身份訊息的10-Breed Dogs數據集，並使用Stable Diffusion將數據集中的每張影像，皆額外生成10張與流浪犬相似的影像以擴增數據集，且建立為SD 10-Breed Dogs數據集。使用上述兩個數據集，分別建立用於品種分類的10BD模型與SD 10BD模型，以及用於身份辨識的10BD ID模型與SD 10BD ID模型。 於犬隻品種分類的實驗結果中，無論哪一個模型的準確率皆可達致95%以上，而身份辨識的結果，則證實了使用包含原始乾淨影像與生成影像進行訓練的SD 10BD ID模型，其性能相較僅使用原始乾淨影像訓練的10BD ID模型更佳。亦即若模型於訓練時已包含寵物犬成為流浪犬之生成影像，即可增加辨識的準確度，進而應用於找尋飼主遺失之愛犬，並藉此改善流浪犬問題。
並列摘要
With the rapid development of cities, the increasing number of people keeping pets, and the adoption of the policy of "zero euthanasia", the issue of stray dogs has gained more and more attention. Additionally, the loss of pets is a major concern and source of anxiety for owners, making dog identification one of the recent popular research topics. This paper proposes a dog breed and identity recognition model based on Xception, utilizing image recognition technology to locate lost dogs. The framework first classifies the breed of dog images and then performs identification within the breed. A 10-Breed Dogs dataset was created for this study, consisting of ten dog breeds with identity information. Moreover, Stable Diffusion was used to generate 10 additional images resembling stray dogs for each image, thus expanding the dataset, which was established as the SD 10-Breed Dogs dataset. Two models were developed using the above datasets: the 10BD model and SD 10BD model for dog breed classification, and the 10BD ID model and SD 10BD ID model for dog identity recognition. The experiments conducted for dog breed classification demonstrated that the accuracy of any model exceeded 95%. The results of identity recognition confirmed that the SD 10BD ID model, which included both original images and generated images for training, outperformed the 10BD ID model trained solely on raw images. In other words, if the model incorporates generated images of pet dogs transformed into stray dogs during training, the recognition accuracy can be improved. This, in turn, can be applied by pet owners to locate their lost dogs and thereby help address the issue of stray dogs.