魚類在我們的日常生活佔了很大一部分,但是在沒有專業的訓練之下,我們很難去 辨別魚的種類。因此,我們需要研發一個魚類影像分割和辨識系統來幫助一般人辨別一 個圖片中可見的每隻魚的種類。不過從頭研發一個上述的系統相當的消耗人力和時間。 本研究提出一個基於深度學習的魚類影像切割和辨識的系統設計,使用 MIAT 系統設計 方法論,按照 IDEF0 進行模組和階層式系統設計,並使用 GRAFCET 進行離散事件建 模,以進行系統軟體高階合成來建構我們的系統。我們亦展示我們使用的影像標註工具 以及標註流程和規則。本研究以一個魚類圖片資料庫來驗證此一方法論建構出的系統。 實驗結果顯示,本系統可以在驗證資料集上面達到 85%的 top-1 準確度,並在兩個月之 內把本系統研發完成。 ;Fish plays an important role in our life, but we can hardly to recognize the type of fish without professional training. As a result, we would like to design a system which can segment the part of fish from an image then classify the kind of each part of segments for ordinary human being. Nevertheless, designing and implementing such system from scratch costs a lot of human labor and time.
This dissertation proposes a fish image segmentation and classification system design with the help of deep learning. We adopt the core concepts of MIAT Methodology to construct the system with IDEF0 for modular and hierarchical system design and GRAFCET of discrete event modeling. Also, we demonstrate the image annotation tool we use on labeling dataset and state the protocols of image annotation. We adopt a fish image dataset to verify the system created with applying MIAT Methodology within two months, and the system shows a top-1 accuracy of 85%.