本研究應用影像紋理分析原理,計算甘藍穴盤種苗影像之均勻度、熵、最大機率對比、k階反向差衝量、相關度、行長機率、同質度、及群聚傾向度等九項紋理特徵,並結合倒傳遞類神經網路,發展辨識甘藍種苗不同生長階段之法則,以自動辨識甘藍種苗一片本葉到五片本葉等五個不同生長階段。經分別試驗比較原始資料轉換方式、取樣影像大小與灰階、共生矩陣計算方向及樣本平均等有關影像處理之因子,及隱藏層數、節點數及訓練次數等類神經網路結構因子,獲知原始資料處理以經移位化處理可得較佳收斂值及辨識率,較佳的影像取樣大小及樣本灰階在類神經網路訓練後之收斂值與較差者比較相差五倍以上,顯示此二因子影響辨識率甚大。網路結構中以具二層隱藏層,節點數分別為7、5者較適用,同時訓練次數不可太多以免造成過度訓練降低辨識度。辨識法則之最佳組合因子為採用32灰階之128×128影像為紋理分析樣本,共生矩陣計算方向為對角線方向,所求得之特徵值經移位化及平均處理者,在網路訓練次數為10萬次,對於甘藍種苗生長階段之正確辨識率可達88.9%,而主要辨識誤差多發生在穴盤苗較茂密的第四葉與第五葉階段。
Image texture analysis was employed in this study for automatic recognition of growth stage of head cabbage seedlings in a nursery tray. Nine textural features such as uniformity, entropy, maximum probability, inverse different moment of order k, correlation, probability of a run length, homogeneity, cluster tendency were determined and fed to back propagation neural network to predict seedling growth stage from one leaf to five leaves. Factors affecting the accuracy of the recognition algorithm were examined. They were input data transformation, sample image resolution and gray scale, coocurrence matrix direction operator, sample image averaging, neural network structure and training epochs. The tests revealed that data transformation with offsetting and averaging, higher sampled image resolution and gray scale yielded better network training convergence. Several network structures were also tested and compared. Among them, network structure with two hidden layers of 7-5 nodes showed better performance. However, over-training should be avoid to secure the recongnition accuracy. The best combination for the recogition algorithm was tested and selected. With sampled image resolution of 128×128, 32 gray-scale, diagonal position operator for the co-occurrence matrix, 100 thousand training epochs with two hidden layers of 7-5 nodes, 88.9% recognition rate was achieved in identifying the seedling growth stages when the nursery trays were densely covered with seedling leaves.