蝴蝶影像辨識基於深度學習及其遊戲化行動裝置之應用

臺灣生態非常豐富，有「蝴蝶王國」之稱，臺灣蝴蝶的數量約400種，其中約有50種是臺灣特有種。民眾在戶外郊遊時，看到了翩翩起舞的蝴蝶，會想得知眼前這些漂亮的蝴蝶是甚麼種類。近年機器學習迅速地發展，可以透過大量數據進行深度學習，自動提取特徵，省去人工分析的時間，辨識速度迅速且準確度高。所以本論文利用深度學習訓練出一套能夠辨識蝴蝶翅膀紋理的模組，並且開發一支 iOS 蝴蝶影像辨識行動應用程式。本論文以臺灣常見的94種蝴蝶，每種蝴蝶各有10張影像，共計940張蝴蝶影像當作資料庫，在 Caffe的框架中進行深度學習，並且使用 AlexNet、GoogLeNet 和 ResNet 三種不同的卷積神經網路，進行實驗比較。此外，本論文提出將蝴蝶影像做翅膀的正規化處理，並且透過新的數據增強 (Data Augmentation) 方式來增加訓練數據到 4700 張影像，希望能夠提高辨識的準確率。實驗結果顯示使用資料增強、蝴蝶翅膀正規化與 GoogLeNet 訓練出來的辨識模型，有最高的 99.912% 辨識率。使用訓練資料集外的 40 張測試影像，亦有90% 的測試準確率。現在是智慧型裝置的時代，因此本論文實作出一個蝴蝶影像辨識的iOS手機應用程式。將辨識率最高的Caffe模型，透過Core ML Tools轉換成Core ML格式，結合Xcode開發手機應用程式，讓民眾能夠透過這個App方便且迅速知道蝴蝶的種類。另外，為了增加民眾使用這款應用程式的意願，我們還在App中增加遊戲化與圖鑑搜尋的功能，希望民眾除了使用辨識功能外，還能從中獲得樂趣，並且認識到蝴蝶生態的相關知識。

關鍵字

深度學習；卷積神經網路；影像辨識；數據增強；智慧型裝置

並列摘要

Taiwan’s ecology is very rich and is known as the "Butterfly Kingdom". There are about 400 species of butterflies in Taiwan, of which about 50 species are endemic to Taiwan. When people are going outdoors, they see dancing butterflies, and they want to know what kind of beautiful butterflies are in front of them. In recent years, machine learning has developed rapidly. Deep learning can be carried out through a large amount of data, features can be automatically extracted, and the time of manual analysis can be saved. The recognition speed is fast and the accuracy is high. This thesis uses deep learning to train a model that can recognize the texture of butterfly wings, and develops an iOS mobile app. This thesis uses 94 species of butterflies that are common in Taiwan, each with 10 images, a total of 940 butterfly images as a database. Perform deep learning in the framework of Caffe, and use three different convolutional neural networks, AlexNet, GoogLeNet, and ResNet, for experimental comparison. In addition, this thesis proposes to normalize wings of butterfly images, and a new data augmentation to increase the dataset to 4700 images, hoping to improve the accuracy of recognition. Experimental results show that the highest 99.912% recognition rate is achieved by using data augmentation, butterfly wing normalization, and GoogLeNet. There is also a 90% test accuracy rate when using 40 test images outside of the training dataset. Now is the era of smart devices, so this thesis has actually implemented an iOS mobile app for butterfly image recognition. Convert the Caffe model with the highest recognition rate into Core ML format through Core ML Tools, then develop the mobile application with Xcode. It lets the public know the species of butterflies conveniently and quickly through this app. In order to increase people’s willingness to use this app, we have also added gamification and pictorial search functions to the app. We hope that people can have fun from it in addition to using the identification function, and learn the knowledge of butterfly ecology.