基于卷积核滤波器筛选的CNN模型精简方法

摘要

近年来,卷积神经网络(CNN)在计算机视觉领域中取得了令人瞩目的成绩,在各类的图像竞赛中取得了可喜的成绩.然而,CNN带来的高精度和鲁棒性的背后是计算量大幅增加的支撑,复杂的深层卷积神经网络往往需要在计算机集群或是高端GPU才能运行,因此CNN很难运行在嵌入式设备中,尤其是运行在手持设备中.这就导致CNN不能从实验室进入到人们的日常生活中.本文提出了一种基于卷积核滤波器筛选策略的CNN模型精简方法.通过分析CNN在前向传播中各神经元的激活情况,来找出对网络模型贡献度高的卷积核滤波器,并将这些滤波器重新封装成一个新的“小CNN模型”.这个小模型在不仅在识别率上拥有很高的性能,而且还有效减低了模型体积和计算时间,在本文中通过实验表明CNN模型能够通过精简的方式使运算速度显著加速,而准确率仅仅只下降了两个百分点.%In recent years,convolution neural networks(CNN)have made remarkable achievement in various computer vision tasks from image recognition to object detection,which shows promising results for many computer vision tasks.However,with highly precise accuracy and robustness,CNNs often require high intensive computation,e.g.,some CNNs are running on clusters of PCs or GPU.Hence,it is difficult to integrate CNNs into consumable products,especially into mobile devices,which leaves a huge gap between laboratory and practical usage in CNN based application.In this paper,we propose a novel simplification method.By analysing the activation of neurons during the feed-forward,finding parts of filters which contribute greatly to some specific tasks and reorganizing the architecture,a smaller scale CNN is obtained.The new CNN not only has high classification or detection accuracy,but also has smaller volume and significantly reduced computational time.Experiments on the several databases demonstrate that the smaller scale CNN model accelerate the speed dramatically,but merely has two percentage loss in classification accuracy.