結合類神經網路及改良式鏈碼特徵抽取技術於自動目標系統

Abstract

我們提出一個結合類神經網路及改良型鏈碼(chain code)特徵抽取技術的 自動目標辨識系統。我們將各種未遭遮蔽的目標物以及它們常遇到的遮蔽 情況(例如:目標物遭碉堡、 建築物、...等遮蔽) 建入資料庫(data base library)，使得目標物在遭受遮蔽時仍可獲得高辨識率。整個系統的架構 包含前置處理(preprocessing)，特徵抽取(feature extraction) 以及分 類器(classifier) 三個部份。前置處理包括影像分割及解決目標物旋轉 問題，使得此系統具有旋轉不變性。特徵抽取採用改良型鏈碼特徵抽取技 術。改良型鏈碼特徵主要是從物體局部邊界抽取更細部特徵藉以提高系統 辨識率及對遮蔽物的容忍度。由於改良型鏈碼特徵具有對目標物平移及大 小的不變性；因此整個系統具有旋轉，平移及大小的不變性。分類器採用 supervised fuzzy adaptive Hamming net，該神經網路具有線上學習的 功能。在實驗中，此系統具有不錯的辨識率且對遮蔽有不錯的容忍度。 We propose an automatic target recognition system that combines a neural net-work and an improved chain-code feature-extraction technique.The system's database library includes targets that are not occluded and targets with commonocclusions(such as blockhouses, buildings, etc.). The system is composed of three parts, preprocessor, feature extractor, and classifier. The preprocessorperforms image segmentation and makes the system results rotation invariant.The feature extractor uses an improved chain-code feature-extraction techniquethat extracts more detailed features from the target's local boundaries in order to increase the system's recognition rate.Because the extracted featuresare position and scale invariant, so the entire system is rotation, position, and scale invariant. The classifier uses a supervised fuzzy adaptive Hamming net that is able to learn on line. In experiments, the system achieved a good recognition rate and proved robust to occlusion.