论文部分内容阅读
具有平移、比例和旋转不变性（TSRI）的模式识别是经常遇到的一个三阶问题。近年来的研究表明：把模式的不变性构建于三阶高阶神经网络（HONN）中是实现TSRI模式识别的有效途径。但对于一幅N×N的图象，三阶HONN为存贮连接权所需的存贮容量正比于N6，这一要求限制了HONN在大尺寸图象中的应用。为解决这问题，我们先用边缘探测和对数螺线映射处理图象，把三阶问题转化成二阶问题，使HONN的存贮需求降至O（N4），再改进二阶HONN的结构，使这一需求进一步降至O（N2）。我们用128×128的图象进行了仿真实验，结果表明：该方法对大尺寸目标图象的TSRI识别切实可行。
Pattern recognition with translation, scaling, and rotation invariance (TSRI) is a commonly encountered third-order problem. Recent researches show that it is an effective way to realize the pattern recognition of TSRI by constructing the invariance of the mode in the third order higher order neural network (HONN). However, for an N × N image, the storage capacity required by third-order HONNs to store connection weights is proportional to N6, a requirement that limits the use of HONN in large-size images. To solve this problem, we first use edge detection and logarithmic spiral mapping to transform the third-order problem into the second-order problem and reduce the storage requirement of HONN to O (N4), and then improve the structure of second-order HONN , To further reduce this demand to O (N2). We use 128 × 128 image simulation experiments, the results show that: This method of large-size target image TSRI recognition feasible.