影像辨識系統在台灣常見蝴蝶種類辨識之應用

Abstract

昆蟲因為特殊的生態習性以及取得容易等特性，常在科普教育中被作為教學素材。 台灣素有蝴蝶王國的美名，極富多樣性的蝴蝶，是台灣最具代表性的昆蟲。在利用昆蟲尤其是蝴蝶作為科普教育材料時，對蝴蝶種類的鑑定，是其中最初步也是最重要的一環。台灣有約四百種之豐富蝶種，對非昆蟲專業人士或熟悉蝶類的相關從事人員，蝶種的鑑定並非易事。在蝴蝶種類辨識上，一般學生及民眾多半利用圖鑑等科普書籍，藉比對書籍上之圖片以及觀察到之蝴蝶進行比對及鑑定，但此鑑定動作需要一定的經驗與知識基礎，難免讓初學者失去興趣，造成利用蝴蝶作為科普材料的障礙。為克服此障礙，本研究嘗試利用影像識別技術，應用至台灣常見蝴蝶種類之鑑定。影像辨識系統中，蝶種的選定，我們設定為台灣的常見蝴蝶。在本研究中，常見蝴蝶之定義為:一般學生或大眾「可見」、「易見」、「想見」之蝶種。並利用科博館已鑑定完成之蝴蝶標本為基礎建立影像，作為提供影像資訊基礎，用以建立比對的模組影像。分析從野外拍攝之自然生態影像以及模組影像，諸如顏色、斑紋等影像資訊，並以影像辨識技術比對兩種影像的相似程度。本研究分析影像在HSV （hue、saturation 及 value）及Lab （Lightness，a 和 b 表示顏色對立維度）色彩屬性空間中的參數平均值以及各參數的比例，用以比對影像在顏色上的相似程度。此外，並使用SURF （Speeded Up Robost Features， 加速穩健特徵）運算比對方法，比對影像在斑紋上的相似程度。並設計一系列測試，用以修改比對方法之設定，找尋效果最佳之比對方法。期望最終此研究成果能應用於智慧型手機或其他攝影通訊器材，使用者可經由上傳影像資訊，在雲端進行影像辨識，增加蝴蝶資訊取得的便利性，進而提高以蝴蝶作為科學教育材料的使用價值。

The availability of 400 species of butterflies in Taiwan is an example of high species richness that is opening opportunities for great science education, but it also could be a challenge of being a material of science education. The difficulties during species identification may create misunderstanding for learning or seeking further information. People usually identify insects comparing the similarity between insects, pictures and descriptions on illustrations. However, these inconveniences, requirements of high knowledge and experiences could left many people behind the knowledge or lose their interest on insects. To breakthrough these barriers, developing a new innovative system that could automatically identify insect species is the one solution. To make that possible, first we define the common butterfly species in Taiwan as the butterflies that have stable population, easy to obtain, or famous species. By photograph, well-identified specimens in NMNS, that butterfly was taken as a test insect and a database images of common butterflies in Taiwan was established. The application of the image recognition techniques to analyze the database images and the pictures captured in natural environment, finding the similarities and differences for in the reorganization of butterfly species were accomplished. The identification characteristics included are colors, patterns and the information link with images captured as time and location. The HSV color model was applied and compared with its histogram, and SURF was used to identify the specific pattern of butterfly. To correct the identification method into most effective way, series of tests were designed by recognizing the texting images which were modeling from ranges of situations in natural images, and the results are expected to be used to adjust the parameters. This work aimed to develop a system that can be broadly applied in cell phones or other photographable communication equipment to enhance the knowledge that ultimately improve the usage value of insect as science education material.