以梯度方向及類神經網路為基礎之頂角偵測系統研究

• Main Author: 吳昇龍
• Other Authors: 陳萬清
• Format: Others
• Language: zh-TW
• Published: 2000
• Online Access: http://ndltd.ncl.edu.tw/handle/23581969429855283127

碩士 === 中正理工學院 === 電子工程研究所 === 88 === In this thesis two gradient-based corner detection approaches, the comparison and the neural network, are proposed to investigate a fast and reliable corner detector. In the comparison approach, the gradient direction map of an image is formed by using the 3x3 and 7x7 Sobel operators. An edge-threshold is adopted to considerably decrease the processing complexity. Since the variations of the gradient directions near the corner are greater than those along the edges, corner detection then can be performed by comparing the direction degree of the target pixel with those who have the largest and the smallest degrees around the target. Comparing algorithms and voting processes are proposed. Corner detection results are presented to verify the reliability performance of the gradient-based comparison approach. A neural network corner detection approach is proposed. Training patterns constructed by eighteen sorted direction degrees around the target pixel are fed into a 18-4-1 neural network and the back-propagation algorithm is adopted during training. Testing results of real images show its robustness performance. Comparisons have been made with the median-based and the SUSAN detectors. Testing examples of the scaled, rotated, various gray-level, and noise-corrupted images are performed to investigate the detection performance. Corner detection results show the superiority and reliability of the proposed gradient-based approaches.