Matching the Conjugate Relations between Two Feature Group Through the Hopfield-Tank Neural Network

• Main Authors: Jin-Jon Tzen, 鄭景中
• Other Authors: Yi-Hsing Tseng
• Format: Others
• Language: zh-TW
• Published: 1994
• Online Access: http://ndltd.ncl.edu.tw/handle/14932045742700734739

碩士 === 國立成功大學 === 測量工程學系 === 82 === In this paper, a reliable and efficient pattern recognition system is developed to find all the conjugate relations between two groups of features. This system conceptually mime the human recognition process.It is expected to be adaptive to disturbances of distortion, noises and differences of orientation between two conjugate features. The system is achieved through the considerations of some quantitative information derived based on the similarity of shape and consistency of orientation. An optimal conjugate relation between two feature groups can be determined after taking all of the quantitative information into consideration. In order to obtain quantified recognition information,features are described by using Fourier descriptors. Then, the shape similarity and the orientation differences are calculated by using the least-squares approach to matching Fourier descriptors. The Hopfield-Tank neural network is used to combine all information and determine the optimal conjugate relation. The neural network simulates the operation of human neuron so as to achieve the flexibility and adaptation of human brain. In the network, the conjugate relations between two feature groups are searched based on an energy value which is a function of the shape similiarity, the orientation consistency, constraints of conjugate relations. The optimal state of the conjugate relations between two groups of features is reached when the minimum energy value is obtained. Features fetched from automatic image segmentation process of two conjugate aerial images and digitized from conjugate aerial images are took as the experiment data to test the feasibility of this recognition theory. The test shows an encouraging result. end.