1 内容介绍
考虑到异常检测问题中正负样本严重失衡,难以满足卷积神经网络训练对样本的要求,提出了基于AlexNet模型的异常检测模型.在数据预处理阶段,通过隔帧采样的方式生成3组训练数据,并利用预训练的AlexNet模型提取相应的3组图像特征,最后通过并联的形式训练3组一类支持向量机模型1SVM,在测试阶段对3个1SVM的结果进行投票,获得最终的检测结果.以UMN数据集作为实验数据进行实验,算法的等错误率为1.8,优于其他算法,充分说明了算法的有效性.
2 部分代码
%% Applying Deeplearning to Anomaly Detection for manufacturing product
% This is the way to detect feature outlier with AlexNet and 1-class SVM kernel method.
clear; close all; imtool close all; clc;rng('default')
% unzip('data.zip')
% winopen('testimage')
%% Read Pre-trained Convolutional Neural Network (CNN) %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
convnet = alexnet() %
%% show layers
convnet.Layers % show layer
%% open folder including training images
rootFolder = pwd;
categ = {fullfile('data','trainingimage')};
winopen(fullfile('data','trainingimage'))
%% use imageDatastore object for dealing with huge amount of image.
imds = imageDatastore(fullfile(rootFolder, categ), 'LabelSource', 'foldernames')
imds.ReadFcn = @(filename) readAndPreproc(filename); % set function to resize image to 227*227*3.
tbl = countEachLabel(imds) % Show the number of training image
%% Run AlexNet to get the feature data at the fc7 layer
fLayer = 'fc7';
trainingFeatures = activations(convnet, imds, fLayer, ...
'MiniBatchSize', 32, 'OutputAs', 'columns'); % run the network with images and get the feature data at the defined layer
%% train a 1-class SVM with the feature data
W = ones(size(trainingFeatures', 1), 1);
d = fitcsvm(trainingFeatures', W, 'KernelScale', 'auto', 'Standardize', false, 'OutlierFraction', 0.04,'KernelFunction','gaussian');
%% Detect 4 abnormal images from test image set
categ2 = {fullfile('data','testimage')};
% Read 100 images as a test set
imds2 = imageDatastore(fullfile(rootFolder, categ2), 'LabelSource', 'foldernames','IncludeSubfolders',true)
imds2.ReadFcn = @(filename) readAndPreproc(filename);
tic % start timer
testFeatures = activations(convnet, imds2, fLayer, ...
'MiniBatchSize', 32, 'OutputAs', 'columns'); % Execute Alexnet and get data at the fc7 layer
[~, score] = predict(d, testFeatures'); % predict score with trained SVM
[score_sorted, idx] = sort(score); % sort by score (is score is small (like negative), the image can be abnormal)
idx(1:25) % the indices of Top 25 abnormal images
toc % Stop time and show the calculation time
%% show the sorted images side-by-side
im = readall(imds2);
im = im(idx); % sort images by score in ascending order
sz = size(im{1});
% Insert rectangle on images people defined as anomaly
for i=1:numel(idx)
if idx(i) <5
im{i} = insertShape(uint8(im{i}),'rectangle',[1 1 sz(1) sz(2)],'LineWidth' ,10);
end
end
I = cat(4, im{1:100});
figure,montage(I, 'Size', [10 10]) % show 10*10 images in a figure
% The score of images in the first row are low. (anomalousness is high)
% the 1-4 lowest score images have rectangle yellow frame.
% This means that prediction by classifier is same as the correct answer people define.
score(idx) %
%% Use t-SNE for visualization
rng default %
testLabels = imds2.Labels; % Use label for visualization
% Use t-SNE to visualize 4096 dimension data bidimensionally
Y = tsne(testFeatures','Algorithm','exact','NumPCAComponents',50,'Perplexity',45);
figure
gscatter(Y(:,1),Y(:,2),testLabels)
title('Default Figure')
% feature plots of abnormal image are located far from center of whole distribution
% classifier detects these outliers
3 运行结果
4 参考文献
[1]付青、罗文浪、吕敬祥. 基于AlexNet和支持向量机相结合的卫星遥感影像土地利用变化检测[J]. 激光与光电子学进展, 2020, 57(17):9.
[2]雷丽莹, 陈华华. 基于AlexNet的视频异常检测技术[J]. 杭州电子科技大学学报(自然科学版), 2018.