【EASRC分类】基于EASRC和ELM算法实现音频脑电图 (EEG) 数据分类附matlab代码

原创已于 2022-08-02 16:31:04 修改 · 222 阅读

0 ·

CC 4.0 BY-SA版权

文章标签：

#matlab #算法 #分类

于 2022-07-29 14:54:24 首次发布

图像处理专栏收录该内容

872 篇文章

订阅专栏

1 内容介绍

最近的研究表明，在图像分类领域，极限学习机（ELM）的速度优势和稀疏表示分类（SRC）的准确率优势。然而，这两种方法都有各自的缺点，例如，通常已知 ELM 对噪声的鲁棒性较差，而 SRC 则很耗时。因此，ELM 和 SRC 在计算复杂度和分类精度上相得益彰。为了统一这种互补性，从而进一步提高分类性能，我们提出了一种高效的混合分类器，以利用本文中 ELM 和 SRC 的优势。更准确地说，所提出的分类器由两个阶段组成：首先，通过监督学习训练 ELM 网络。其次，采用关于获得的 ELM 输出的可靠性的判别标准来决定查询图像是否可以正确分类。如果输出可靠，则由 ELM 进行分类；否则查询图像将被提供给 SRC。同时，在 SRC 阶段，通过 ELM 输出提取与查询图像而不是整个字典相适应的子字典。因此可以减少 SRC 的计算负担。手写数字分类、地标识别和人脸识别的大量实验表明，所提出的混合分类器在分类精度上优于 ELM 和 SRC，具有出色的计算效率。

2 仿真代码

function nn = elm_initialization(nn)

% biases and input weights
scale = 1;

nn.b = 2*scale*rand(nn.hiddensize,1,'single')-scale;
nn.W = 2*scale*rand(nn.hiddensize, nn.inputsize,'single')-scale;

% nn.b = randn(nn.hiddensize,1);
% nn.W = randn(nn.hiddensize, nn.inputsize);

% if nn.orthogonal
% if nn.hiddensize > nn.inputsize
% nn.W = orth(nn.W);
% else
% nn.W = orth(nn.W')';
% end
% nn.b=orth(nn.b);
% end

end

% ========================================================================= -------------------------------------------------------------

clear;clc; addpath(genpath('./.')); rng('default');

%%%%%%%%%-----Load data--------------------------------------------------

load(fullfile('data','AR.mat'));

traindata = traindata./( repmat(sqrt(sum(traindata.*traindata)), [size(traindata,1),1]) ); testdata = testdata./(repmat(sqrt(sum(testdata.*testdata)), [size(testdata,1),1]) );

% [traindata,PS] = mapminmax(traindata,-1,1);% % testdata = mapminmax('apply',testdata,PS);

% [traindata,PS] = mapstd(traindata);% % testdata = mapstd('apply',testdata,PS);

[trainlabel1, testlabel1] = label_convert(trainlabel, testlabel,'2');

%-----------Setting---------------------------------------------------- alpha = 0.1; kclass = length(unique(trainlabel))/2; % for adaptive class domain selection nn.hiddensize = 1000; % hidden nodes number

method = {'ELM','RELM'}; nn.activefunction = 's'; nn.inputsize = size(traindata,1); nn.method = method{2}; nn.type = 'classification'; %-----------Initializzation-----------

nn = elm_initialization(nn);

%--------RELM-LOO-------------------- nn.method = method{2}; nn.C = exp(-4:0.2:4); [nn, acc_train] = elm_train(traindata, trainlabel1, nn); %[nn, acc_test] = elm_test(testdata, testlabel, nn);

ID = []; IDe = []; lamda = 5e-4; tol = 1e-2; tic; f = 0; j=1;

%--------RELM-LOO---SRC-------------------- for i = 1 : size(testdata,2) [nn, acc_test] = elm_test(testdata(:,i), [], nn); O = nn.testlabel; [Tf, id] = max(O); O(id) = -inf; [Ts, id2] = max(O); Tdiff = Tf-Ts; IDe = [IDe, id]; if Tdiff > alpha ID = [ID, id]; % if id ~= testlabel(i) % fprintf('Wrong classification for %1.0f th testing sample by ELM criterion (|T_first - T_second| = %1.2f) \n ', i, Tdiff); % end else f = f + 1; [sim, slabel] = sort(nn.testlabel, 'descend'); newtrainlabel = trainlabel(ismember(trainlabel,slabel(1:kclass))); newtraindata = traindata(:,ismember(trainlabel,slabel(1:kclass))); y = testdata(:,i); s = l1_ls(newtraindata, y, lamda, tol, 1); newlabel = unique(newtrainlabel); % if ~ismember(testlabel(i), newlabel) % fprintf('Wrong classification for %1.0f th testing sample by ELM criterion (|Adaptive class domain|) \n ', i); % end gap = []; for indClass = 1 : length(newlabel) coef_c = s(newtrainlabel==newlabel(indClass)); Dc = newtraindata(:,newtrainlabel==newlabel(indClass)); gap(indClass) = norm(y-Dc*coef_c)^2; end wgap3 = gap ; index3 = find(wgap3==min(wgap3)); id3 = index3(1); id = newlabel(id3); fprintf('%1.0f / %1.0f %1.0f %1.3f %1.0f %1.2f %1.0f \n', i, size(testdata,2), f, sum(s), find(slabel==testlabel(i)), Tdiff, id==testlabel(i)); ID = [ID, id]; end end

Rec_ELM = sum(IDe==testlabel)/length(testlabel); Rec_EASRC = sum(ID==testlabel)/length(testlabel); % recognition rate

disp(['RELM_LOO recogniton rate is ' num2str(Rec_ELM)]);

disp(['EASRC recogniton rate is ' num2str(Rec_EASRC)]);

figure plot(testlabel,'bo'); hold on plot(IDe,'g+'); plot(ID,'r*'); legend('真实值','RELM','EASRC') title(['RELM_LOO recogniton rate is ' num2str(Rec_ELM),',EASRC recogniton rate is ' num2str(Rec_EASRC)])