以CNN進行植物圖片的辨識以及經過處理後的再辨識

Abstract

本研究的目標為利用卷積神經網路來辨識照片中的植物所歸屬的類別。在模型架構方面採用了知名的VGG-16模型，並使用遷移式學習（transfer learning）來降低學習至參數收斂所需要的時間。本次使用的資料集為擁有500種品種、約10萬張圖片的植物照片資料集。為了對植物特徵或主體較不明顯的圖片更進一步的處理，在一般的辨識過程後，加入了Unseen Category Query Identification辨識法來選取出這些辨識度不足的圖片，之後對選取出的圖片進行圖像分割，嘗試將植物主體與背景或雜訊分離，藉此來強調植物特徵。執行完前述步驟之後，選出包含植物主體的圖片，將其輸入模型進行再辨識，並實行其它實驗來比較這些做法的成效。

In this work, we want to recognize the species of plants in a picture by using Convolutional Neural Networks (CNN). We use the VGG-16 model in our experiments. To make the training process converged efficiently, we train model by leveraging transfer learning. The dataset we use is made up of 500 species consisting of approximate 100,000 plant images. We employ Unseen Category Query Identification (UCQI) after the prediction step and picking those images which don't have obvious features or main bodies. For those picked images, image segmentation is used for separating plant from other objects and background noise. We choose the segmented images containing plant main body for re-classification. Detail comparisons between the proposed method and baselines are shown on the experimental part.