本論文提出一多模態階層式分類器應用於大類別的植物種子辨識，我們首先擷取種子影像的不同模態特徵，包括LBP紋理、HOG特徵、GLCM特徵、BLOB幾何特徵、色彩統計特徵等五種模態特徵。每一種模態各自使用一個非監督的自組織映射神經網路進行分群，將特徵相近的種子聚類為同一群，不同的模態特徵可能有不同的分群結果。我們規劃了一個投票機制用以從五個模態分群結果篩選出模糊類別，做為第一階層的初分類結果。接著我們使用每一類種子的特徵映射聚類座標，以One-hot編碼方式，將五組座標合成一組二值座標向量作為新特徵向量，建立每一類種子的機率神經網路分類模型，進行第二個階層的植物種子辨識。我們以多模態階層式分類器對311類種子樣本進行實驗，最高能達到86%的辨識率，而以TOP-5為輸出辨識率更可高達95%。與深度神經網路中的ResNet50進行比較，結果可高出24%，運算時間與所使用資源量也遠少於深度神經網路。;This paper proposes a multi-modal hierarchical classifier for the identification of large-scale plant seeds. We first capture different modal features of seed images, including LBP texture, HOG features, GLCM features, BLOB geometric features, color statistical features, etc. Five types of modal features. Each modality uses an unsupervised self-organizing map neural network for clustering seeds with similar features into the same group, and different modality features may have different clustering results. We have planned a voting mechanism to screen out fuzzy categories from the five modal grouping results as the first classification results of the first class. Then we use the feature mapping cluster coordinates of each type of seeds, and use the one-hot encoding method to synthesize five sets of coordinates into a set of binary coordinate vectors as new feature vectors to establish a probabilistic neural network classification model for each type of seed. Carry out the second class of plant seed identification. We use multi-modal hierarchical classifier experiments on 311 types of seed samples, and the highest recognition rate can reach 86%. The recognition rate can be as high as 95% with TOP-5 as the output. Compared with ResNet50 in the deep neural network, the result can be 24% higher. The calculation time and the amount of resources used are far less than the deep neural network.
