- 學位論文
運用校正板與鏡頭光學參數的內視鏡影像校正法
A Distortion Correction Method for Endoscopy Images Based on Calibration Patterns and Parameters of Optical Lens
摘要
近三十年來，隨著光學工程、資訊與電子技術的進展，內視鏡已成為內、外科醫學不可或缺的寶貴工具。由於其臨床應用呈指數成長，故已成為一個專科領域。為了在狹小的管道中取得大面積的影像，內視鏡大多使用廣角鏡頭（魚眼鏡頭），因此所取進來的影像都會產生一定程度的扭曲形變失真。此種扭曲情況以中心點為基準，呈現輻射狀向外擴散並愈趨嚴重的趨勢。本論文探討此影像扭曲的效應及其修正，利用簡單的校正板與數學校正模型，對一組內視鏡儀器進行形變影像校正。只需對同一台內視鏡進行一次校正，日後便可重複利用所得到的數學模型直接校正內視鏡影像。 校正板經由內視鏡取像之後，我們必須利用數位影像處理技術將校正板圓點從扭曲的影像中萃取出來。我們提出一個二階的數學模型，並考量廣角鏡頭的光學參數，將萃取出來的校正板圓點給定座標之後，代入數學模型中來進行內視鏡影像校正。實驗結果顯示我們提出的方法是有效的，因為在與原始校正板影像的比較之下，失真影像與校正後影像相對區域面積計算的平均誤差分別為76.46%與4.68%。此外，我們也利用類神經網路來對校正板的圓點座標進行訓練，可以達到不錯的校正結果，其中隱藏層所含節點的個數也會影響到校正結果。
並列摘要
In the last 30 years, the progresses in optical engineering, computer science and electronic techniques have made the endoscopy an invaluable tool in both internal clinics and surgical operations. As its applications increase exponentially, it has even become a specialized division in the clinical medicine. In order to obtain a larger field of view inside a small and narrow pipeline, the endoscope is usually equipped with wide-angle lens (Fish eye lens). Thus, the acquired images are often with certain degrees of shape distortion. The distortion gets even more serious as the objects extend outward from the center of the lens in radial. This thesis discusses as the effect of image distortion and the correction of the effect. By using a calibration pattern, the nonlinear distortion is corrected with a simple mathematic model for the endoscopic images. Once the endoscopic lens is calibrated, the same mathematic model parameters can be utilized repeatedly for the endoscopic images. After capturing the calibration pattern using an endoscopic instrument, we must apply digital image processing techniques to extract the calibration pattern from the distorted image. We propose a second order mathematic model and consider the parameters of optical lens. The coordinates of each dot in the calibration pattern are the input to the mathematic model for the correction of endoscopic images. The experimental results show that the correction method is effective because by comparing to the original calibration pattern image, the average area calculation errors are 76.46% and 4.68% for the distorted and corrected images, respectively. In addition, we use artificial neural network to train the dot coordinates in the calibration pattern and the training result is good. The number of nodes in the hidden layer can affect the correction result.