基於適地性之即時蝴蝶辨識系統

以往蝴蝶的辨識只能透過先行以相機拍照，並委託專業人士經由人工方式逐一檢視拍攝到的照片，再比對可能的蝴蝶種類，耗費的人力及資源相當龐大。蝴蝶辨識上會受到環境樣貌、蝴蝶花紋及狀態、蝴蝶大小等狀況影響，而過去使用機器學習作法辨識蝴蝶結果較不好，因此本論文採用Single Shot Multibox Detector的深度學習架構辨識蝴蝶種類，並且結合手機程式和辨識模型，協助觀光導覽、蝴蝶復育人員快速辨識蝴蝶種類。本系統以兩個核心架構所組成，第一部分為蝴蝶物件辨識系統，用來找出圖片內的蝴蝶種類及位置，為了驗證模型辨識的正確率，將圖片切分為9份的訓練圖片、及一份的訓練圖片，10次訓練的TOP-1平均辨識正確率為87.85%；第二部分為開發即時的適地性蝴蝶辨識系統，透過GPS判斷使用者位置後，比對該地點擁有的常見、特色蝴蝶類型，達成即時且具適地性的蝴蝶影像辨識系統，改善以往需要投入大批人力辨識蝴蝶種類的問題，並降低辨識模型耗用的手機資源。

關鍵字

Tensorflow ； Single Shot Multibox Detector ；多目標物件辨識；影像辨識；適地性系統

並列摘要

In the previous time, the identification of butterflies can only be implemented by examining the taken pictures and photos one by one via the eyes of experts then compared with the possible types and varieties of butterflies which will inevitably consume a lot of manpower and resources. The recognition of butterflies will be affected by environmental appearance, butterfly pattern and state, butterfly size, etc. Due to the reason that the use of machine learning to identify butterflies is unpleasant, therefore this research undertakes the profound learning structure of the Single Shot Multibox Detector to identify the butterfly types and combine mobile programs with identification models to facilitate tourism guides and help rehabilitators to recognize the butterflies more efficient. The system is consisted of two core architecture. The first part is the butterfly object recognition system which is used to locate the types and location of the butterflies on the picture. The picture is divided into nine training pictures, and one training picture, the average recognition of TOP-1 accuracy for ten trainings is 87.85%. The second part is to develop a location-based real-time application of butterfly recognition. After determining the user's position through GPS, the butterfly image recognition system can be achieved compared with the common and characteristic butterfly types owned by the location which improves the large demand of manpower in the past, and reduce the amount of mobile phone resources used by the identification model.