訓練自己的人臉檢測分類器（級聯+LBP的Matlab的實現）
阿新 • • 發佈：2019-01-05
本文通過MATLAB實現，能夠實時檢測識別到人臉，與OpenCV模型檔案相容，版本最好matlab2017a及其以上，老版本沒試過。本文主要分為3個步驟：（1）攝像頭獲取人臉正樣本影象；（2）攝像頭獲取負樣本影象；（3）訓練識別部分，可選擇從圖片，視訊，攝像頭實時識別。
注意事項：
（a）其中變數isSample=1時，即首次執行需要採集人臉影象，以後請把isSample置為0，表示以後不需要採集正樣本；(b)負樣本產生我寫在另一個函式createNegativeImgs（）裡面，大家執行它即可，負樣本一定不要有你自己的人臉影象哦~ (c)importdata（）函式用於把正樣本的標記檔案匯入到MATLAB工作空間中，請注意格式。
正樣本可以自己手動標記人臉框，可以從trainingImageLabeler APP互動工具獲得，當然方便起見，我 從已有的人臉檢測器xml檔案檢測人臉，從而直接得到人臉正樣本，當然可以匯入到trainingImageLabeler 檢視預覽（注意格式），我這裡直接用的是lbpcascade_frontalface.xml分類器。
直接上程式碼，如下；
%% 用xml預訓練的分類器對人臉進行篩選，記錄人臉，用於訓練,測試 cam = webcam();% 攝像頭介面，沒有的話從matlab central網站搜尋下載 %% 收集樣本 isSample = 0; %這裡如果現場從攝像頭獲取你的影象作為訓練樣本，請把該值置為1 if isSample==1 fig = figure; axes('parent',fig) detector = vision.CascadeObjectDetector('lbpcascade_frontalface.xml'); detector.MinSize = [110,110]; videoPlayer = vision.VideoPlayer; % 人臉檢測與標記 if ~exist('images','file') %當前目錄是否存在images資料夾，沒有則新建 mkdir images end fid = fopen('images/face_rect.txt','a');% 以追加的方式進行寫入 while ishandle(fig) filename = [cd,'/images/',datestr(now,'yyyy-mm-dd-HH-MM-SS-FFF'),'.png']; frame = snapshot(cam); bbox = step(detector,frame); imwrite(frame,filename); fprintf(fid,'%s %5d%5d%5d%5d \r\n',filename,bbox); if isempty(bbox) fprintf(fid,'\r\n'); end positions = bbox; nums = size(positions,1); strLabels = {'face'};%strEye = repmat({'eye'},1,nums-1); RGB = insertObjectAnnotation(frame,'rectangle',positions,strLabels,'color','g'); step(videoPlayer,RGB); end fclose(fid); end %% 不需要訓練 facerect1 = importdata(); imageNames = cellstr(facerect1.imagenames); rects = [facerect1.x,facerect1.y,facerect1.w,facerect1.h]; faceRect = table(imageNames,rects,'VariableNames',{'imageFilename','face'}); index = ~isnan(rects(:,1)); faceTrain = faceRect(index,:); % faceRect.imageNames = cellstr(imageNames); % faceRect.rects = rects;%mat2cell(rects,ones(1,length(labels.imageNames))); num = length(faceTrain.imageFilename); %% 正樣本製作 trainPosNums = 500; % 這裡設定你的訓練正樣本數量，根據你的樣本量適當選擇 newTrainLabels = faceTrain(randi(num,1,trainPosNums),:); %table型別 %% 負樣本製作 trainNegNums = 500; % 這裡設定你的訓練負樣本數量，根據你的樣本量適當選擇 negativeImgDataStore = imageDatastore(fullfile(cd,'NegativeImgs')); negNUM = length(negativeImgDataStore.Files); negativeImages = negativeImgDataStore.Files( randi(negNUM,1,trainNegNums) ); %% 開始訓練 xmlName = 'myLBPfaceDetector.xml'; trainCascadeObjectDetector(xmlName,newTrainLabels,negativeImages,... 'FalseAlarmRate',0.1,'NumCascadeStages',20,... 'FeatureType','LBP'); %% test ,選擇跑的內容 detector = vision.CascadeObjectDetector(xmlName); detector.MinSize = [100 ,100]; detector.MergeThreshold = 4; videoPlayer = vision.VideoPlayer; %% flag選擇平臺，flag = 0為跑圖片，flag = 1為跑視訊檔案，flag=2為跑攝像頭 flag = 2;% 選擇 index = 0; if flag == 0 %跑圖片 imdsTest = imageDatastore('F:\video\patform_data\6月\06',... 'includeSubfolder',true);%圖片檔案，這裡設定你自己的測試人臉影象路徑 for i = 1:length(imdsTest.Files) imageTest = readimage(imdsTest,i); bbox = step(detector,imageTest); RGB = insertObjectAnnotation(imageTest,'rectangle',bbox,'face'); step(videoPlayer,RGB); index = index+1; disp(index); end elseif flag == 1 % 跑視訊 obj = vision.VideoFileReader('F:\video\smokeVideo2017_3_1\170405151456_1280328332795.mp4');%注意這裡是你自己的視訊檔案路徑 while ~isDone(obj) frame = step(obj); bbox = step(detector,frame); if ~empty(bbox) RGB = insertObjectAnnotation(frame,'rectangle',bbox,'face'); else RGB = frame; end step(videoPlayer,RGB); index = index+1; disp(index); end elseif flag == 2 % 跑攝像頭 while 1 % command Window按ctrl+c終止迴圈 frame = snapshot(cam); bbox = step(detector,frame); RGB = insertObjectAnnotation(frame,'rectangle',bbox,'face'); step(videoPlayer,RGB); end else disp('your input may be wrong!'); end
另外importdata()函式和createNegativeImgs()函式如下：
function faceRect = importdata() %% Initialize variables. filename = 'E:\MATLAB\trainMyCascadeFace\images\face_rect.txt'; delimiter = ' '; %% Format for each line of text: % column1: text (%s) % column2: double (%f) % column3: double (%f) % column4: double (%f) % column5: double (%f) % For more information, see the TEXTSCAN documentation. formatSpec = '%s%f%f%f%f%[^\n\r]'; %% Open the text file. fileID = fopen(filename,'r'); %% Read columns of data according to the format. % This call is based on the structure of the file used to generate this % code. If an error occurs for a different file, try regenerating the code % from the Import Tool. dataArray = textscan(fileID, formatSpec, 'Delimiter', delimiter, 'MultipleDelimsAsOne', true, 'TextType', 'string', 'EmptyValue', NaN, 'ReturnOnError', false); %% Close the text file. fclose(fileID); %% Post processing for unimportable data. % No unimportable data rules were applied during the import, so no post % processing code is included. To generate code which works for % unimportable data, select unimportable cells in a file and regenerate the % script. %% Create output variable facerect1 = table(dataArray{1:end-1}, 'VariableNames', {'imagenames','x','y','w','h'}); %% Clear temporary variables clearvars filename delimiter formatSpec fileID dataArray ans; faceRect = facerect1;
-------------------------------------------------分割線-------------------------------------------
function createNegativeImgs()
cam = webcam();
if ~exist('NegativeImgs','file')
mkdir NegativeImgs
end
videoPlayer = vision.VideoPlayer();
index = 0;
while 1
filename = [cd,'/NegativeImgs/',datestr(now,'yyyy-mm-dd-HH-MM-SS-FFF'),'.png'];
frame = snapshot(cam);
imwrite(frame,filename);
step(videoPlayer,frame);
index = index+1;
disp(index);
end RGB = insertShape(frame,'FilledRectangle',bbox,'Opacity',1,'color','red');
