Design Of Convolutional Neural Networks In Tiny Image Classification

Since human society entered information age,varies of kinds of information has been generated in a faster and faster way.In order to dig valuable information from image data by computers,it's a fundamental requirement of increasing the accuracy of image classification.However,traditional image processing methods are not capable of the huge amount of image data.Luckily,with the help of powerful CPUs and GPUs,deep convolutional neural networks make a breakthrough in this problem.In this thesis,convolutional neural network is used for tiny image classification.Two scenes of image data are analysed,which are handwritten digits and natural images.First,the handwritten digits recognition problem is tackled.A simple yet effective convolutional neural network CNN-1 is designed on the basis of LeNet-5,which decreased the error rate by 0.32% compared with the latter by optimization.In addition,the traing speed of CNN-1 is faster than LeNet-5.Then,to deal with the more complex natural images,a deeper convolutional neural network CNN-2 is proposed and examined,which utilized plenty of deep learning tricks.The experiment results show the fact that CNN-2 increased the accuracy by 5.76% compared with the original model and 19.16% compared with CNN-1.What' more,CNN-2 has a light weight structure.This thesis shows the process of optimization of the model through a lot of experiments.It can be concluded from the results that the two convolutional neural networks designed for the two scenes of tiny images are efficient.