一种计算二维图象矩的神经网络方法

A Neural Net Approach to Computation of Two Dimensional Image Moments

提出了一种计算二维图象矩的神经网络方法。首先，导出一维信号矩与一维Ｈａｄａｍａｒｄ变换之间的关系。通过２Ｎ个一维Ｈａｄａｍａｒｄ变换和极少量加法、移位及乘法运算可计算二维图象矩。然后给出用Ｔａｎｋ－Ｈｏｐｆｉｅｌｄ神经网络实现Ｈａｄａｍａｒｄ变换的方法，并对该网络作了改进。改进后的网络呈模块式结构，仅有一个互连强度矩阵且神经元数目减少一半。本文提出的Ｈａｄａｍａｒｄ变换神经网络用变换矩阵的转置直接作为互连矩阵，而无学习步骤，可同时处理整个数据向量，并在几百纳秒内任意逼近所求变换值。该网络适合于实时计算二维图象矩，可进一步用于图象或目标识别。

A neural net approach to computation of two dimensional image moments is presented. First, the relationship between one dimensional signal moments and one dimensional Hadamard transform (1D HT) is derived. Two dimensional image moments can be computed through 2N 1D HT′s and a small amount of addition, shift and multiplication operations. Then, a Tank Hopfield neural net is presented for the computation of Hadamard transform. Some modification of the neural net is described. The modified neural net achieves structural modularity. It has only one interconnect conductance matrix and neural elements are reduced to half. The proposed Hadamard transform neural net uses the transposed transform matrix as the interconnect conductance matrix and has no learning phase. It operates on entire data vectors simultaneously and settles into a value arbitrarily close to the correct transform value within hundreds of nanoseconds. The neural net is suited for computing two dimensional image moments in real time and it is further applicable to image or object recognition.