- 期刊
- OpenAccess
以機器視覺與類神經網路分級玫瑰切花之研究
Cut Roses Grading with Machine Vision and Meural Network
摘要
本文主要目的在使用機器視覺發展分級玫瑰切花的影像處理技術，再以類神經網路來學習人工的分級經驗，建立一套結合機器視覺技術及類神經網路的玫瑰切花分級系統。 本研究對每一枝切花攝取兩張彩色影像，一為整枝切花影像供花莖外觀特徵的擷取，另一為花苞部分影像做為花苞特徵的分析之用。花莖分割的方法為先定義出花莖影像形態集合，再依據此集合對整張影像進行逐行搜尋，以標記出花莖部分。玫瑰花苞的分割則使用彩色影像分割方法再加上膨脹收縮技術進行，完整的保留了玫瑰花苞的色彩資訊。對每枝切花共擷取10個外觀特徵參數，其中代表花莖彎曲情形的有最大彎曲角度、最大偏移距離及平均偏移距離，代表花莖粗細的有底端直徑、中間直徑及頂端直徑，而代表花苞開度的則有投影面積、周長、細密度及長短軸。以花莖彎曲情形、花莖粗細及花苞開度做為分級參數，使用誤差倒傳遞類神經網路來模擬切花外觀品質的人工分級作業；長度分級部分則以影像處理程式直接進行。 本研究長度分級的正確率為93%，實驗所得的最佳類神經網路含一層隱藏層，輸入參數三個，其辨識率為70.7%。
並列摘要
The objective of this study is to develop digital image processing techniques to extract feature parameters of cut roses, and to use the neural network to simulate the manual grading experiences for cut roses grading. Two color images were grabbed for each rose, one of which was the whole cut rose image for analyzing the morphological features of the stem, the other one was the bud image for analyzing the bud features. To segment the stems, the stem image characteristics were defined first, then the image was searched column by column based on the defined characteristics, and finally the stem segments were labeled. To segment the bud image, the color segmentation and the dilation and erosion techniques were utilized and the color information of the bud was not changed. Ten feature parameters were extracted for each cut rose. The stem straightness parameters were the maximum crooked angle, the maximum deviated distance, and the average deviated distance. The stem diameter parameters were the bottom diameter, the middle diameter, and the top diameter. And the bud maturity parameters were the projected area of the bud, the perimeter, the compactness, and the principal axes. Part of the 10 features were selected and inputted to an error back-propagation neural network to simulate human quality grading operations for cut roses. The length grading was run only by the image processing program. The cut roses length grading accuracy is 93%, and the identification rate with the best neural network model obtained in this study is 70.7%, comparedwith human grading results.