改良式粒子群演算法及其在影像接合上的應用

Abstract

現今照相機在工業視覺檢測扮演很重要的角色。各種電子零件越來越精密，為了能清楚觀測物體的細節，通常會調高照相機的解析度。但由於照相機的像素是固定的，調高解析度同時卻也減小了可視的範圍，雖然可以更換較高像素的照相裝置，但是如此一來成本勢必提高，而且像素提升也有其極限，為了改善可見視野縮小的問題，本論文擬提出一個無接縫之影像接合方法，將多張來源影像接合成單張全景影像，以此達到提高影像的解析度卻又不失去可見範圍的目的。硬體方面計畫將照相裝置架設在一個x,y平台上，此平台可以沿著水平與垂直方向移動，如此一來便能對物體拍攝多張不同位置的影像 本論文提出一個影像接合演算法，利用粒子群演算法搜尋最佳接合位置。希望藉由粒子群演算法快速收斂之特性，達成有效率地搜尋接合位置之目的。傳統粒子群演算法可能出現收斂到區域最佳解的情形，因此我們提出之改良式粒子群演算法，加入了一種可以跳出區域最佳解的突變機制，使得陷入區域最佳解時能有機會能重新收斂至全域最佳解，使得搜尋結果更為精確。此外，我們將影像接合結果交由一個最終的確認階段，此階段的用途在於判斷接合結果的好壞，若是判斷為錯誤接合結果，則重新搜尋接合位置，若判斷為正確則進入影像融合的階段，影像融合階段能將因影像亮度不均勻所產生的邊界給消除。在最後的影像接合實驗之中，分別統計標準粒子群演算法及我們提出之改良式粒子群演算法之接合準確率結果，我們發現使用改良式之結果，能將正確率提升22.3%，此外最終確認階段又可將整體正確率提升20.2%左右，因此我們的系統是非常可靠的，且平均產生一個接合結果只需1.75秒，在效率上也有很好的表現。

In the application of industry, cameras play an important role in the visual inspection, the objects are getting more and more precise recently. In order to observe the objects details, it is needed to heighten the resolution of camera. Because the pixels of camera are fixed, heightening the resolution will reduce the field of view at the same time. Reducing the field of view will make the system unpracticed, we could solve this problem by changing high pixels camera, but it is costly and there still have a limit of high pixels camera. In this thesis a seamless image stitching system is proposed to solve reducing field of view problem by stitching all pictures into one panoramic view. We can get both resolution and field of view by this procedure. The hardware in this thesis is a platform with an x, y direction movable axis, and set a camera on the axis. It can take multiple pictures of different regions。 This thesis tends to accomplish a fast image stitching algorithm by utilizing particle swarm optimization algorithm. Because it’s fast convergent property, it can search stitching position efficiently. The problem of basic particle swarm optimization is may converging at local best value, so we modified particle swarm optimization algorithm by adding a mutation scheme that can escape from local best value. There have more chance to converge at global best value, hence the modified version outperform than original one. The stitching result is sent to a final verification scheme. The design of the scheme is to measure the goodness of image stitching result. If the stitching result is rejected by final verification scheme, modified PSO will search another stitching result again. If the stitching result is accepted, image fusion stage will be processed. The target of image fusion stage is to eliminate the boundary cased by illuminant difference of source images. We take an experiment on image stitching. In our five test image sets, the modified version of PSO can improve accuracy rate about 22.3%, furthermore, the final verification scheme with restart PSO can improve accuracy rate about 20.2%, The image stitching algorithm proposed in this thesis has been successfully evaluated that the processing average time is 1.75 second per stitching result, and it is quite efficient.