基于仿生识别系统的多尺度解决方法

Solutions to Multi-scale Problem Based on the Bio-inspired Recognition System

基于神经网络实现的仿生识别系统不仅存在输入尺度限制问题,还由于数据格式的差异不兼容现有解决方案.针对此问题,提出用于仿生识别系统的3种不同多尺度输入解决方法,分别为基于输入事件的缩放法,基于动态窗口的多尺度池化法以及一种新的基于池化层特征的特征缩放法.实验基于相同的前馈分类系统,进行各方法的资源消耗和系统识别率对比.结果表明,多尺度池化法对应的权值数仅为其他两种方法的3.83%,但识别率较低;相比事件缩放法,所提出的特征缩放法能够提升识别率5.54%,算法执行次数减少59.16%,适用于仿生识别系统.

Bio-inspired recognition system based on neural network limits the size of images.The existing methods are not suitable for this system because of its special data format.Aiming at this problem,three different methods are proposed for bio-inspired recognition system,including the scaling method based on input events,the multi-size pooling method based on dynamic windows,and a new scaling method based on features of pooling layers.Based on the same feedforward recognition system,experiments are carried on to make contrasts on the resource consumptions and the recognition rates between different methods.The results show that the total number of weights used in the multi-size pooling method is only 5.54%of other methods,but the multi-size pooling method has the worst recognition performance;In comparison with the event scaling method,the feature scaling method proposed can achieve 5.54%higher recognition rate and 59.16%less number of operations,which is more suitable for bio-inspired recognition system.