應用圖形辨識於青萍數量之計算

Using Pattern Recognition in Counting Numbers of L. Aequinoctialis Welwitsch

蕭友晉(Yo-Jin Shiau) ； 張文亮(Wen-Lian Chang)

《農業工程學報》 53卷1期 (2007/03) Pp. 13-21

https://doi.org/10.29974/JTAE.200703.0002

摘要

本研究之目的在檢測圖形辨識科技辨識青萍數量的能力。研究選用台北縣新海橋人工濕地的青萍作為樣本，並採用圖形辨識科技來計算限定區域水面上的青萍數量。實驗建立了五種不同的浮萍覆蓋密度，以比較本科技計算青萍數的準確度。研究假設，浮萍葉片主要可分為亮綠色及暗綠色，並採取亮暗色階各－的方式分別建立兩組高斯函數來進行辨識工作。
在無葉片重疊的條件下，經由混合高斯分佈模式，五組覆蓋密度下的浮萍計算準確度分別為，95.7%, 76.8%, 68.2%, 82.1%, and 90%。
在兩組高斯分佈混合以及90%可信賴區間的基礎下，浮萍計算的準確度將可達到80%。

關鍵字

圖形辨識 ； 混合高斯分佈 ； 青萍

並列摘要

The purpose of the study is to examine the ability of pattern classification technology for identifying the quantity of duckweed. The study chose the duckweed from Xin-Hai-Qiao constructed wetlands in Taipei County as the study sample, and it applied the pattern classification technology to calculate the quantity of duckweed on top of water surface in a limited area. In this experiment, five covering densities were built to compare the duckweed-counting accuracies of pattern classification technology. The assumption is that duckweed leaves are mainly light green and dark green. Then light pixels and dark pixels were adopted to build two groups of Mixture Gaussian Models for duckweed identification.
While using simplified Mixture Gaussian Model, without overlapping of duckweed leaves, the counting accuracies under five covering conditions were 95.7%, 76.8%, 68.2%, 82.1%, and 90%.
On the bases of two Gaussians and 90% acceptance region, the duckweed counting accuracy will reach 80%.

並列關鍵字

Pattern recognition ； Gaussian distribution ； L. Aequinoctialis