影像辨識一直以來都是機器學習的主流之一，傳統機器辨識與訓練，多採取人工特徵擷取法，然而運用此方法要找出強健的特徵去使模型學習處理數十百萬的資料是不容易的，隨著深度學習、卷積神經網路的開發，文字辨識已運用在日常生活各領域，但科學家對此機器的辨識速度與精準度的要求也不斷提高。 本研究採用卷積神經網路程式為研究範例，透過田口品質工程的理念，尋找影響其品質特性之控制因子與變動水準數，製作L18直交表制定實驗順序，搭配Google Colaboratory軟體執行取得數據，依據實驗結果計算其S/N比，繪製出最佳化控制因子反應圖，其中以輸出層激勵函數採用Softmax，池化層為Pool_size(2，2)，卷積層為Tanh，拋棄層為0.8，隱藏層激勵函數採用Elu，隱藏層神經元為256，訓練方式優化器為Adam，訓練方式準確率模型為Binary_Accuracy為最佳組合，最後進行確認實驗驗證結果，以最佳化組合平均值為0.9971，S/N比為-0.025，優於其他組合結果，確認該組合為最佳組合，為以供爾後研究人員做為參考。 Image recognition has been a mainstream topic in machine learning. Conventionally, machine recognition and training has been conducted through manual feature extraction, which requires completing the challenging task of finding robust features for models to learn to read millions of data. Following the introduction of deep learning and convolutional neural network, text recognition technology has been integrated in daily living. However, the scientific demand for the speed and accuracy of machine recognition continues to increase. With technology evolving continually, mass data integration and application poses a major challenge in machine learning, in which the selection of parameters in algorithms is critical. This study will focus on the convolutional neural network program. Through Taguchi quality engineering, the control factors to and levels of changes in the quality of a convolutional neural network will be explored. An L18 orthogonal array will be established to formulate the experiment procedure. Data will be collected using Google Colaboratory. Signal-to-noise ratio and quality characteristics will be calculated according to the experiment results, and the effects of various factors will be explained. The experiment will be repeated with the parameters of the control factors adjusted to optimize the factor set.
