MATLAB卷积神经网络——基于ResNet-50进行图像分类

news2024/11/22 1:50:42

一、ResNet50工具箱安装

（1）下载工具箱

https://ww2.mathworks.cn/matlabcentral/fileexchange/64626-deep-learning-toolbox-model-for-resnet-50-network

（2）在matlab打开下载的resnet50.mlpkginstall文件

（3）使用下面代码进行测试，出现结果说明安装成功


clear
clc
% Access the trained model
net = resnet50();
% See details of the architecture
net.Layers
% Read the image to classify
I = imread('peppers.png');
% Adjust size of the image
sz = net.Layers(1).InputSize;
I = I(1:sz(1),1:sz(2),1:sz(3));
% Classify the image using Resnet-50
label = classify(net, I);
% Show the image and the classification results
figure
imshow(I)
text(10,20,char(label),'Color','white')

二、训练猫狗数据集

（1）数据集下载链接：

https://pan.quark.cn/s/e043408353a5

（2）将数据集按照如下目录进行放置

（3）生成预训练模型

在命令行窗口输入 deepNetworkDesigner(resnet50)

然后点击导出→使用初始参数生成代码

保存生成的网络初始化参数，生成的mlx文件可以叉掉：

修改文件路径，类别数目以及相关参数：

clear
clc
filename = "datasets"; 
%% 加载用于网络初始化的参数。对于迁移学习，网络初始化参数是初始预训练网络的参数。
trainingSetup = load("resnet-50.mat");
%% 设置图像文件夹路径和标签
nc = 2;  %类别
imdsTrain = imageDatastore(filename,"IncludeSubfolders",true,"LabelSource","foldernames");
[imdsTrain, imdsValidation] = splitEachLabel(imdsTrain,0.8);  % 80的训练集
%% 调整图像大小以匹配网络输入层。
augimdsTrain = augmentedImageDatastore([224 224 3],imdsTrain);
augimdsValidation = augmentedImageDatastore([224 224 3],imdsValidation);
%%  设置训练选项
opts = trainingOptions("sgdm",...
    "ExecutionEnvironment","gpu",...
    "InitialLearnRate",0.01,...
    "MaxEpochs",20,...
    "MiniBatchSize",64,...
    "Shuffle","every-epoch",...
    "Plots","training-progress",...
    "ValidationData",augimdsValidation);

三、训练及测试结果

（1）训练结果

（2）导入一张图片进行测试

clear
clc
load result\net.mat
load result\traininfo.mat
%% 随便选一张进行测试
[file,path] = uigetfile('*.jpg');
if isequal(file,0)
   disp('User selected Cancel');
else
   filename = fullfile(path,file);
end
I = imread(filename);
I = imresize(I, [224 224]);
[YPred,probs] = classify(net,I);
imshow(I)
label = YPred;
title(string(label) + ", " + num2str(100*max(probs),3) + "%");