模擬退火參數偵測系統於物件偵測與震測圖形之應用(II)

Abstract

類神經網路的理論及應用之研究，近年來在國際上愈來愈重要，應用範圍也相當的 廣泛。今年現在做的國科會計畫，我們提出利用模擬退火演算法的圖形偵測系統，偵測 直線、圓、橢圓、與雙曲線的參數。不同於赫夫轉換類神經網路(Hough transform neural network)會受到收斂到區域最小值的影響，此一系統利用模擬退火演算法求得的參數， 會使得誤差為全域的最佳化(最小值)，因而能夠使得偵測的直線、圓、橢圓、與雙曲線 的參數將更精確，但受限於圖形的公式，不能同時偵測橢圓與雙曲線，所以我們提出新 的圖形公式及同時式的圖形偵測系統，改進目前的系統，使其可同時偵測橢圓與雙曲 線。我們也提出階層式的圖形偵測系統，可減少大量參數之計算，及可用於大量圖形之 偵測上。此兩系統皆能夠改善傳統赫夫轉換(Hough transform)需要大量記憶體空間的缺 點，而階層式的圖形偵測系統因其初始溫度與降溫次數較少，能比同時式的圖形偵測系 統，得到更快的收斂。本計畫第一年我們將作同時式的圖形偵測系統，第二年我們將作 階層式的圖形偵測系統。在實驗上，此兩系統於影像中大量圖形偵測成功後，我們將應 用於偵測震測圖形中的直線的直接波與雙曲線的反射波的參數，其震測圖形包含模擬震 測圖形與實際震測圖形，偵測的結果將有助於震測訊號的解釋與進一步的處理。

Recently the development of theory and application of neural networks become increasingly important in international community. And the areas of the applications are quite wide spread. In this year (2007), our project is supported by the National Science Council. We propose the pattern detection system using simulated annealing that can detect the parameters of the lines, circles, ellipses, and hyperbolas. The system can not only detect parameters with a higher precision, but also a globally optimal solution rather than a locally optimal solution by the Hough transform neural network. But it can not synchronously detect ellipses and hyperbolas due to the limitation of the used formulas of patterns. So we propose the synchronous system to detect ellipses and hyperbolas. We also propose the hierarchical system to reduce the number of parameters in computation, and to detect the large number of patterns. Traditionally, Hough transform needed the large memory space, but the two proposed systems of simulated annealing can reduce the requirement of large memory. In the first year we use the synchronous system to detect patterns. In the next year, we use the hierarchical system to detect patterns. After the success of both systems in image pattern detection, we will apply it to detect the parameters of the line of direct wave and the hyperbola of reflection wave in the simulated one-shot seismogram and real seismic data. The detection results of both systems will improve the seismic interpretations and the further seismic data processing.