本研究旨在探討深度學習模型”多層感知神經網路”及”卷積神經網路”在圖像辨識上的訓練結果比較，也研讀當今常用的優化算法，熟知損失函數對參數更新的影響關係，兩個模型訓練將以蝴蝶圖片進行實作。 經由線上開放的圖片庫網站，取得9141張共五類蝴蝶，並自製成數據樣本集，分別帶入兩個深度模型，透過自行建立隱藏層結構，觀察兩者訓練時間及訓練準確率，在迭代結果上分析擬合情形。而後再進一步引入PCA降維方法，對數據預處理，看看對於圖片背景降維效果能否提高訓練或驗證準確度。 ;The goal of this thesis is to explore the training results of two deep learning models "multilayer perceptual neural network" and "convolution neural network" in image recognition, and study the popular optimization algorithms. To this topic, we study the influence of loss functions on iterative parameters, and demonstrate two models with the pictures of butterflies. There are five types of butterflies with 9141 pictures obtained through the online database website. We use these pictures for the files of data samples to two deep learning models, and observe the training time and accuracy by establishing the hidden layers. Then, we analyze the fitting situation to the iterative results. Finally, by using principle component analysis(PCA) in dimension reduction method, we preprocess the data and observe the reduction effect of the image background so that we can improve the training or verification accuracy.
