應用深度學習於垂直綠牆果實之影像辨識

Abstract

近年來台北市努力推動垂直綠牆和綠屋頂等政策來降低城市熱島效應、改善城市景觀與淨化城市空氣。在2007 年台北市公部門開始建議重大工程的工地圍籬要有垂直綠牆。美國、歐洲與新加坡漸漸地開始在垂直綠牆面種植蔬果類植物，使垂直綠牆變成垂直農場的一種。此研究與德國柏林工業大學合作，設計一台在垂直綠牆的自動採收機，在設計自動採收機第一步是要使採收機有機器視覺，能辨識蔬果種類與定位，因此本研究開發了一個程式，可以使用在垂直綠牆上進行百香果的定位、數目的計算及 面積的計算等。此程式架構主要為兩大步驟，第一步驟為收集四百多張百香果圖片、並加以標註，再使用 YOLO v3 深度學習模型進行訓練，得到百香果的自動化辨識、初步定位和數目的計算。第二步驟為使用高斯模糊、HSV 色彩轉換、遮罩運算等影像處理方式來做到更進一部準確的果實中心點定位、果實大小計算與輪廓的偵測。此程式以人機介面的形式呈現，讓使用者在使用上更佳的容易上手，平均準確度為 97.93%。

In recent years, Taipei City Government has worked hard to promote vertical green walls and green roofs to reduce heat island effects, improve urban landscapes and purify urban air. In 2007, the public sector in Taipei city began to propose vertical walls for the construction site fence. The United States, Europe, and Singapore have gradually begun to grow fruits and vegetables on vertical green walls, turning vertical green walls into vertical farms. This research cooperates with the Technical University of Berlin in Germany to design an automatic harvesting machine in the vertical green wall. The first step in designing the automatic harvesting machine is to make the harvester have machine visual and recognize the type and positioning of fruits and vegetables. A program was developed to use the positioning, number calculation and area calculation of passion fruit on vertical green walls. The program architecture is mainly composed of two major steps. The first step is to collect and label more than 400 images of passion fruit, and then use the YOLO v3 deep learning model to train, and obtain the automatic identification, preliminary positioning and number calculation of passion fruit. The second step is to use image processing methods such as Gaussian blur, HSV color conversion, mask calculation, etc. to achieve more accurate fruit center point, fruit size calculation and contour detection. This program is presented in the form of GUI, which makes it easier for users to use it. The average accuracy of this program is 97.93%.