一种新的复杂细胞启发下的特征提取方法

鉴于V1区复杂细胞具有提取外界刺激不变本质的能力,设计了一个提取目标图像局部不变特征的方法.该方法首先使用提出的无监督算法（PCICA）,从图像中学习出类似于复杂细胞感受野的滤波器集合.然后利用这些滤波器组成的复杂细胞描述子,提取目标图像各个位置的不变特征.最后对图像特征图进行分块统计,将各区域的直方图序列作为图像的最终描述.测试结果表明,PCICA具有类似于快速独立分量分析算法（FastICA）三阶收敛的特点,从图像中学习出的滤波器集合,表现出复杂细胞感受野的拓扑结构.这些滤波器对于局部图像的微小变化并不敏感,对于检测不变特征十分有利,并在MNIST手写体数据库上取得0.84%的识别错误率.

A new complex cell-inspired feature extraction method

Inspired by the ability of complex cells to extract the invariant features from stimuli,a feature extraction method is designed to extract features that are invariant to minor variations of input data.First,the proposed learning algorithm,named Pairwise Cumulant-based Independent Component Analysis（PCICA）,is used for modeling the properties of complex cells in primary visual cortex（V1）.Then,a set of invariant feature maps are extracted by the learnt filters from the each image.Finally,the feature maps are respectively divided into several regions,and each image is described by the histogram sequence produced from all these region patterns.The simulation results show the high efficiency of PCICA.These learnt filters are good at extracting invariant features because they are less sensitive to the small changes of input,and achieve the testing error rate of 0.84% on MNIST database.