機器視覺應用於影像對位已屬於成熟技術，產業界大多採用商業化套裝模組，由軟體或硬體供應商出售專用擷取設備或賣出軟體程式碼與資料庫資及輔助開發服務，廠商各自擁有專門代理權或影像技術，由於處理影像基礎原理存有差異，以致不同軟體系統特性各有優劣，整體來說商業化系統使用者僅就業者所提供之限制功能範圍內開發銜接程式碼，並無從知悉系統核心，且軟體本身屬於線性結構，不能針對因多重影像物件組合造成非線性狀態求最佳化對位補正位置進行處理。本研究收集類神經相關應用文獻，選擇倒傳遞類神經網路進行多腳位元件影像對位最佳化找尋，分析理論原理後將其編寫為程式碼導入自製影像對位系統，並以隨機拍攝三腳位電子元件影像二百張，涵蓋90%腳位座落之整體與個別腳位座標，以平均差統計較佳之學習速率0.1與慣性因子0.1，以此組合教導系統學習最佳化，經過2000次學習與收斂，得到元件樣本影像對位最佳化非線性模擬函數，此函數即網路由經驗訓練而得之加權值與偏權值(閥值)，倒傳遞類神經網路將無法量化與公式化處理的問題以非線性曲線表示，自此不論輸入允許範圍內之元件座標為何，網路皆得以直接進行正向傳遞將結果輸出。廣意言之，倒傳遞類神經網路為一經驗值策略輔助系統。 Image registration are skillful of machine vision, the commerce service package was adopted as the drawing of equipment most of the time , software source code library and database or image capture hardware and research assistant sold by professional vendor , they have unique authority of agency or own vision technique, there are differences as the basis principles of image processing, characteristics of different software systems that have merit and demerit , the system are linear programming could not to process non-linear problem of multi-pin interlace location optimizated .This research method combine Back Propagation Neural Network to image alignment software we design for multi-pin alignment optimization and capture two hundred pieces of sample image with random , that could be mantle over 90% whole and each of pin location , using MSE to find the best speed learn rate 0.1 and momentum coefficient 0.1 . During 2000 times training for BPNN , network showing convergence , than promote bias and weight , these could be drawing curve approach truely , the cruve solve non-linear phenomenon , input any grid in range the Neural network running forward counting and directly to output result with optimization .
