使用機器學習與影像處理反算光彈圓盤受力

Abstract

為了反算光彈圓盤的受力，本研究將藉由光彈圓盤受力後所產生的應力條紋進行影像分析，研究中將使用機器學習(支援向量機、隨機森林樹)以及利用深度學習建立卷積神經網路模型，對於光彈圓盤進行應力條紋光彈力學（Photoelasticity）的應力特性分析。 本研究利用影像處理與機器學習將光彈影像進行濾波，將光彈影像之亮暗場影像進行影像分類與受力反算，實驗結果發現，我們能夠藉由實驗中所拍攝的亮暗場光彈圖片反算出受力間距為1kg的光彈影像；為了提升受力間距的細緻度，我們進一步將亮場圖片和暗場圖片做影像處理，希望獲取更多的條紋特徵，實驗結果顯示，相較先前只能反算出受力間距為1kg的光彈影像，做傅立葉轉換後的影像處理所獲得的條紋倍增光彈影像，在實驗上能更夠有效且細緻的反算出受力間距為0.5kg的影像，因此後續研究都將以條紋倍增的光彈影像進行最精確的受力分析。 研究最後將訓練好的模型，對訓練資料以外的未知光彈影像進行受力反算，並搭配本研究中所提出的三種修正方法提升準確率，研究結果表明，模型針對不同的資料集的預測結果使用判別法修正後，所得到新資料集和原始資料集相比，能夠提升20%的準確率。

In order to determine the force on the photoelastic disk, this study will employ image analysis on the stress fringes generated by the photoelastic disk. In the research, machine learning (support vector machines, random forest) and deep learning will be used. We will establish a convolutional neural network model to analyze the stress characteristics of the photoelastic disk. This study uses image processing and machine learning to filter photoelastic images, classify and determine the types and force of bright and dark field images of photoelastic images. According to the experimental results, we can determine the photoelastic image with a force spacing of 1kg from the bright and dark field photoelastic images in the experiment. In order to improve the detail of the force spacing, we using image processing of bright field pictures and dark field pictures further, in order to obtain more stripe features. The experimental results show that compared with the previous photoelastic image, which could only be determined with a force spacing of 1kg, the stripe multiplied photoelastic image obtained by subtracting the bright field image and the dark field image can be more effective and detailed determine the photoelastic images with the force spacing of 0.5kg. Therefore, in subsequent study, we will use the photoelastic image with stripe multiplication to accurate analysis of the force on images. At the end of the study, the trained model will be used to determine the force on unknown photoelastic images , and combined with the three correction methods which are proposed in this study to improve the accuracy. After correcting the prediction results data set by the correction method, the results show that after empolying the correction methods , the prediction data is more linear fit to the true data set .When we compare the new data to the original data, we can find that the accuracy is increased by 20%.