摘要
具有大小、位移、旋转不变性（TRSI）的目标识别是一个三阶问题,如果采用三阶神经网可以使问题得以解决,但最大难点就是互连权矩阵的组合爆炸问题,在本文中,我们将视觉生物物理学的研究成果应用到高阶神经网的研究中,首次提出双向Log-Polar变换与高阶网（HONN）结合的方法,有效地解决了权系数的组合爆炸问题,进一步提高了识别率。取得了突破性进展。使得将高阶网用于自动目标识别成为可能。
Translational、rotational、scaling invariant (TRSI) pattern recognition is an important problem in the automatic target recognition (ATR) field. But neither traditional pattern recognition algorithms nor neural network approaches have yet provided satisfactory solution for this problem after years of study. Recent research has shown that the high order neural networks (HONN) of order three have numerous advantages over other neural network approaches in respect of the object recognition with invariant of the object's position, size, and in-plane rotation. The major limitation of HONNs is that the number of connected weights is too large to store on most machines. In this paper, we have developed an integrated method which combines the bidirectional log-polar mapping and HONN pattern recongnizer. It reduces the HONN memory requirement from O(N6) to O(N2). The proposed method has been successfully verified. The result shows that this method can indeed achieve TRSI pattern classification. 100% accuracy can be guaranteed for noise-free test image. In addition, its high robustness to noise、pattern deformation and partial occlusion makes it very useful for real world applications.
出处
《中国体视学与图像分析》
1996年第1期19-19,共1页
Chinese Journal of Stereology and Image Analysis
