- 學位論文
使用深度學習於花朵辨識之研究與應用
Study and Implementation of Flower Image Recognition using Deep Learning
摘要
台灣向來以物種繁多出名，有著很多植物，到處可見各式各樣的植物與花朵。在以前要認識植物，只能靠著當場翻閱圖鑑，或是拍照等回到家之後才能查詢。現今科技進步，可以透過智慧型裝置，運用網路查詢關鍵字甚至直接拍照，利用影像直接進行搜尋。像是Google以圖搜圖，但辨識的結果並沒有很好，另外，之前的辨識系統，大多是使用傳統的辨識方法，使用者需輸入影像後，需要經過正規化、分割、擷取特徵，再與資料庫中的特徵做比對，過程繁複且特徵擷取的部分需要人工分析。近年人工智慧發展迅速，而深度學習則是透過大量數據反覆學習，自動擷取特徵，這不但省去人工，而且快速、準確度又高。 本論文以國立暨南國際大學校園內常見的50種植物的花，每種植物各取10張花朵，共計500張影像為資料庫，使用NVIDIA DIGITS訓練，並分別使用 AlexNet和GoogLeNet兩種網路架構，進行三組對比實驗。實驗一：使用原始的、沒有經過任何修改的影像做訓練及測試，AlexNet Top1辨識率64%，GoogLeNet Top1 辨識率64.8%；實驗二：使用去除背景的影像做測試及訓練，AlexNet Top1辨識率57.2%，GoogLeNet Top1 辨識率63%；實驗三：使用去除背景、切除留白部分的影像做測試及訓練，AlexNet Top1辨識率76%，GoogLeNet Top1 辨識率78.8%，實驗三的辨識結果最好，證明影像經過去背正規化，會有較高的辨識率。 除了暨大校園花朵資料庫，也使用ImageCLEF 2013 Plant Identification提供的植物資料庫Pl@ntView中的花朵影像進行實驗，共224種花、3506張影像，model訓練的結果AlexNet Top1 辨識率64.0625%，GoogLeNet Top1 辨識率70.5966%。 最後，本論文實作出一個花朵影像辨識的iOS APP，將DIGITS訓練好的model，透過Core ML Tools轉換成Core ML格式，放入APP中，使用者只要打開APP，利用手機鏡頭拍攝實時影像，這些影像便會在model裡進行辨識，最後將Top1結果顯示在畫面下方，讓使用者可以簡單、方便且快速的得知花朵的種類，以便認識、親近各種花朵。
並列摘要
Taiwan has always been famous for its wide variety of species, especially the plants. Many kinds of plants grow everywhere. In the past, people, with strong eager and curiosity to know the plants they came across, had to look up the plant encyclopedia on the spot, or took pictures first and then went back home to search the information and identification of the plants. Out of the organs of the plant, flower is more eye-catching and pleasing. Nowadays, with the help of advanced technology, people can easily access smart devices, putting in keywords, even the images, to search on the internet. At present, most flower recognition systems follow the traditional recognition method. Users need to input images, after normalization, segmentation, feature extraction, recognition system then matches features from the dataset. The process is complicated and feature extraction needs manual analysis. The Artificial Intelligence (AI) technology develops promptly and promisingly. Amid AI, deep learning systems, repeatedly self-trained by learning large amounts of data, and automatically extract features instead of manual work. Therefore, it becomes more efficient, accurate and without manual interfering. In this thesis, the author employs Deep Learning GPU Training System (DIGITS) and two network architectures, AlexNet and GoogLeNet, to undergo three sets of experiments. The images of fifty kinds of flowers in National Chi Nan University campus, ten images of each kind, totally five hundred images, serve as the database. Experiment (1) is the approach of using original image to train and test, and the recognition accuracy rates of AlexNet and GoogLeNet in Top1 are 64% and 64.8% respectively. Experiment (2) is the approach by removing the background of the image, the recognition accuracy rates are 57.2% and 63%. Experiment (3) is the approach of removing background, cutting out the blank area and focusing the flower to become normalized, the recognition accuracy rates are of 76% and 78.8%. As above mentioned, the Experiment (3) turns out significantly to be the best result. It is proved that the image will have a higher recognition rate after removing background and being normalized. In addition, the author also uses flower images in Pl@ntView, provided by ImageCLEF 2013 Plant identification, to do the experiment same as Experiment (3). There are 224 kinds of plants, including 3506 images in the dataset. Model training results show the recognition accuracy rates are 64.0625% and 70.5966%. Finally, another purpose of this thesis is to develop a convenient application for real time flower recognition, based on smart hand-held devices. The author makes an iOS APP for flower image recognition. Converting the trained Model by DIGITS into Core ML format and putting it into smart phone, a real time flower recognition APP has been implemented in iOS system. Through real time image in the phone, users can recognize flowers via Model and has result in real time. Users simply open the APP to capture real-time images. These images will be recognized in the model, and the Top1 results will be displayed.