蛾類多科別以影像為基礎之自動辨識

在台灣的夜晚隨處可見飛蛾，且飛蛾種類有5000多種，民眾需要花費大量的查詢時間。而民眾會將飛蛾的樣貌拍下來，以影像中的飛蛾外型去判斷相似科別，再比對顏色與紋理進行辨識。本論文改進論文 [1] 提出的蛾類影像辨識系統並新增多科別蛾類便是，使用者能以簡單的操作能快速的得知飛蛾的資訊，查詢結果會以相似度最高的飛蛾影像做排序，能大量的減少一般民眾查詢的時間。本論文以902張的尺蛾科 (Geometridae) 的標本影像做全域顏色特徵之實驗，以Sub-block Color Match擷取影像顏色特徵，特徵比對方法使用 Histogram Intersection、Approximate Nearest Neighbor (ANN) 兩種方法比較準確度與搜尋時間。在全域顏色特徵的實驗中Histogram Intersection的上翅顏色比對Top60的準確度為53.88%、搜尋時間為0.7秒，而 Approximate Nearest Neighbor 上翅顏色比對 Top60 的準確度為 44.77%、搜尋時間為0.022秒，實驗結果可以看出 Histogram Intersection準確度較高但ANN搜尋的時間快上32倍。另外本論文進行多種局部描述子之研究，參考 [7] 將Scale Invariant Feature Transform (SIFT)、Raw Intensity (RI)、 Geometric Blur (GB)、DAISY、Local Intensity Order Pattern (LIOP) 5種單一描述子的特性融合用於比對，而參考論文 [7] 中的比對方法，有Ratio、CAT、Ranking、Density四種多描述子的融合方法。資料集是從資料庫挑出5組飛蛾標本影像，以人工的方式挑出正確的配對用於驗證各種描述子比對效能。在我們實驗結果中可以看出Density效果較佳，但Density需要大量的計算花費時間較長尚未能應用系統，後續研究可以往改善Density的計算時間發展。在多科別實驗裡包含尺蛾科、裳蛾科、夜蛾科、草暝科有1619種標本影像，飛蛾影像為每種1~2張共2823張。將所有的影像作前置處理，之後特徵擷取利用環形顏色直方圖與二維傅立葉描述子進行單翅及雙翅單一特徵與多特徵比對實驗，在雙翅顏色比對Top 60 為33.63 %和紋理比對 Top 60為31.31%，飛蛾雙翅 Linear Evidence Combination 的實驗結果 Top 60 準確度為 38.87%，可以看出多種特徵比單一特徵辨識較佳。

關鍵字

蛾類辨識；環形顏色直方圖； ANN ； Ratio ； CAT ； Ranking ； Density

並列摘要

There are moths everywhere at night in Taiwan, and the species of moths are about 5,000 kinds, so people need a lot of time to query the species of the unknown moth. They will take some pictures from the moth, and use the contour of the moth to classify its family, then recognize by matching color and texture. This thesis improves the image-based moth recognition system [1] by adding multi-families moth recognition. Users can get the moths information rapidly with simple operation. Moths images of query results will be sorted by similarity which can greatly reduce the user’s query time. Our experiment of global color feature applied to the 902 sample images of Geometridae. Sub-block Color Match is used to extract feature, and feature matching methods are Histogram Intersection and Approximate Nearest Neighbor (ANN). Precision and searching time are evaluated. In experiment results of global color feature, color matching of top wing is 53.88% by Histogram Intersection in Top 60 precision and searching time is 0.7 seconds, Approximate Nearest Neighbor is 44.77% in Top 60 precision and searching time is 0.022 seconds. So we can know Histogram Intersection is higher in precision, but ANN is faster 32 times than Histogram Intersection in searching time. In addition, we also study the other experiments about multiple local descriptors. Referring to [7], we apply the multiple descriptors fusion method from 5 local descriptors which are Scale Invariant Feature Transform (SIFT), Raw Intensity (RI), Geometric Blur (GB), DAISY, Local Intensity Order Pattern (LIOP). And refer the paper [7], we know 4 matching methods as Ratio, CAT, Ranking and Density to fuse multiple descriptors. Dataset is 5 pairs of moth sample images, and each pair has correct matches by manual selection, which verify the matching performance from every descriptor. In our experiment results, we can know Density method has the better result, but Density needs a lot of time to compute, so we can't use on our system. Further research to reduce the computing time of Density is necessary. In the experiment of multiple famalies, there are 1,619 species that include Geometridae, Erebidae, Noctuidae and Crambidae. Dataset is 2,823 sample images, and every species includes 1 or 2 images. We do pre-processing on every image, and extract feature by Sub-block-color-Match and 2-D Fourier descriptor to do matching experiments of one wing and two wings that are about matching by single feature and multiple features. In the experiment results, color matching of two wings is 33.63% in Top 60 precision, and texture matching of two wings is 31.31% in Top 60 precision. The experiment result of two wings is 38.87% in Top 60 precision by Linear Evidence Combination. We can know that multiple kinds of feature is better than single kind of feature in moth recognition.