中文摘要：

提出了一种新的特征合并机制,模拟初级视皮层中复杂细胞汇聚合并来自不同简单细胞的响应.该机制首先对简单细胞的线性响应进行归一化,然后将同频率和方向下不同相位简单细胞的响应进行能量合并,最后取局部邻域内响应的最大值作为合并后的输出,得到对输入刺激具有一定相位和平移不变性的不变特征.将其应用于目标识别,在MNIST手写数据库上的测试结果表明：基于新的合并机制的方法能取得更低的识别错误率,对目标的局部变化有更强的鲁棒性.

英文摘要：

A new feature pooling operation was proposed,which cou ld model the scheme of complex cell to pool response from different simple cells in primary visual cortex.First,the new operation normalized the linear respon se of simple cells.Then,the normalized response of quadrature pairs were poole d by the energy model where the quadrature oairs are cells with the same frequen cy and direction but different phases.Finally,the most active response among l ocal areas was selected as the final output,showing some phase and shift invari ance to input stimulus.An object recognition method was proposed based on the n ew feature pooling operation.Experimental results on the MNIST database show th at the method achieves lower error rate,and is more robust to local transformat ions of inputs.