在使用matlab处理数据标签时，

1，首先要判断这张图片标签是否是符合规则的，对于不符合要舍去。在某些情况下，可能不能完全舍去。那么就需要，根据保留标签的特点，来存储数据。比如 使用case 语句。

2，标签可以过多或者需要融合，可以根据原始图片的特点来融合数据。

 

采用MATLAB，对数据进行自带工具箱的使用练习
准备：matlab，Data。其中本人matlab为2016b,工作空间中Data为540行12列，540个样本，11个维度，最后一列为标签。

打开maltab软件，输入“classificationLearner”，回车。
输入数据，选择“From Workspace”.
因为我的样本是行排列，则选择“use columns as variables ”,最后一行数据选择Response。右栏为校验选择。选择好后选择“start session”.
点图为某两个维度之间的显示，可以选择右侧进行观察多个。工具栏选择ALL，注意“use parallel”不选。点击“train”.将得到多个分类器的分类准确度结果，其中最高的会被自动标注。

Label Images for Classification Model Training
matlab—-标签图像分类训练工具
http://cn.mathworks.com/help/vision/ug/label-images-for-classification-model-training.html?s_tid=srchtitle
Description
The Training Image Labeler provides an easy way to label positive samples that the trainCascadeObjectDetector function uses to create a cascade classifier. Using this app, you can:
Interactively specify rectangular regions of interest (ROIs).

交互指定感兴趣的举行区域


Using the ROIs, you can detect objects of interest in target images with the vision.CascadeObjectDetector System object.

借助于ROI用户可以使用`vision.CascadeObjectDetector对象检测目标图像中的感兴趣对象。


You can load multiple images at one time, draw ROIs, and then export the ROI information in the appropriate format for the
trainCascadeObjectDetector. The labeler app supports all image data formats that the
trainCascadeObjectDetector function uses.

用户可以一次加载多幅图片，画出ROI并能够以适当的方式为trainCascadeObjectDetector导出ROI信息。

Open the Training Image Labeler
MATLAB® Toolstrip: Open the Apps tab, under Image Processing and Computer Vision, click the app icon.


打开App工具栏，在Image Processing and Computer Vision子项目下，打开trainingImageLabeler程序。如下图所示：



打开如下图所示： 
MATLAB command prompt: Enter trainingImageLabeler. 

使用命令trainingImageLabeler打开工具。
App Controls
You can add an unlimited number of images to the Data Browser. You can then select, remove, and create ROIs, and save your session. When you are done, you can export the ROI information to an XML file.


用户可以想Data Browser中添加图片，并且图片的数量没有限制。用户可以选择、移除和创建用户感兴趣区域（ROI），并且可以保存用户会话。当用户完成以上工作时，用户可以将ROI信息导出为XML文件。
Add Images
Use the Add Images icon to select and add images to the
Data Browser. You can add more images at any time during your editing session. The source images remain in the folder where they were originally located. The app does not make copies of the original images or move them. If you
 rotate the images in the Data Browser, the app overwrites the images in their original location with the modified orientation.
The app provides a list of image thumbnails that you loaded for the session. Next to each thumbnail, you see the file name and number of ROIs created in that image.

Specify Regions of Interest
After you load images, you can delineate ROIs. You can switch between images and continue creating ROIs. Drag the cursor over the object in the image that you want to identify for an ROI. You can modify the size of an ROI by clicking either the corner or
 side grips. To copy and paste an ROI, left-click within it’s border to select it. You can select one or more ROIs to move or to copy and paste. To delete an ROI, click the red x-box,, in the upper-right
 corner.

You can also use the following shortcuts:
Control-C to copy
Control-V to paste
Control-X to cut
Remove, Rotate, and Sort Images
You can remove, rotate, or sort the images. Right-click any image to access these options. To select multiple images, press
Ctrl+click. To select consecutive images, press Shift+click. To sort images by the number of ROIs, from least amount of ROIs contained in each image, right-click any image and select
Sort list by number of ROIs.
New Session
When you start a new session, you can save the current session before clearing it.
Open Session
You can open a new session to replace or add to the current session. The app loads the selected .MAT session file. To replace your current session, from the Open Session options, select Open an existing session. To combine multiple sessions, select Add session
 to the current session.
Save Session
You can name and save your session to a .MAT file. The default name is LabelingSession. The saved session file contains all the required information to reload the session in the state that you saved it. It contains the paths to the original images, the coordinates
 for the ROI bounding boxes for each image, file names, and logical information to record the state of the files.
Export ROIs
When you click the Export ROIs button, the app exports the ROI information to the MATLAB workspace in a 1-by-M structure, where M represents the number of images. The structure contains two fields. One field stores the image file location and the other field
 stores the corresponding ROI information for each image. You are prompted to name the variable or to accept the default positiveInstances name. The first field, imageFileName, contains the full path and file name of the images. The app does not copy and resave
 images, so the stored path refers to the original image and folder that you loaded the images from. The second field, objectBoundingBoxes, contains the ROI [x, y, width, height] information.

train_data是训练特征数据, train_label是分类标签。 
Predict_label是预测的标签。 

MatLab训练数据, 得到语义标签向量 Scores(概率输出)。

1.逻辑回归(多项式MultiNomial logistic Regression)

Factor = mnrfit(train_data, train_label);
Scores = mnrval(Factor, test_data);

scores是语义向量(概率输出)。对高维特征，吃不消。

2.随机森林分类器（Random Forest）

Factor = TreeBagger(nTree, train_data, train_label);
[Predict_label,Scores] = predict(Factor, test_data);

scores是语义向量(概率输出)。实验中nTree = 500。 

效果好，但是有点慢。2500行数据，耗时400秒。500万行大数据分析，会咋样？准备好一篇小说慢慢阅读吧^_^

3.朴素贝叶斯分类（Naive Bayes）

Factor = NaiveBayes.fit(train_data, train_label);
Scores = posterior(Factor, test_data);
[Scores,Predict_label] = posterior(Factor, test_data);
Predict_label = predict(Factor, test_data);
accuracy = length(find(predict_label == test_label))/length(test_label)*100;

效果不佳。

4.支持向量机SVM分类



Factor = svmtrain(train_data, train_label);
predict_label = svmclassify(Factor, test_data);

不能有语义向量 Scores(概率输出)

支持向量机SVM(Libsvm)



Factor = svmtrain(train_label, train_data, '-b 1');
[predicted_label, accuracy, Scores] = svmpredict(test_label, test_data, Factor, '-b 1');

5.K近邻分类器 （KNN）



predict_label = knnclassify(test_data, train_data,train_label, num_neighbors);
accuracy = length(find(predict_label == test_label))/length(test_label)*100;

不能有语义向量 Scores(概率输出)



IDX = knnsearch(train_data, test_data);
IDX = knnsearch(train_data, test_data, 'K', num_neighbors);
[IDX, Dist] = knnsearch(train_data, test_data, 'K', num_neighbors);

IDX是近邻样本的下标集合，Dist是距离集合。 

自己编写, 实现概率输出 Scores(概率输出)

Matlab 2012新版本:



Factor = ClassificationKNN.fit(train_data, train_label, 'NumNeighbors', num_neighbors);
predict_label = predict(Factor, test_data);
[predict_label, Scores] = predict(Factor, test_data);

6.集成学习器（Ensembles for Boosting, Bagging, or Random Subspace） 

Matlab 2012新版本:



Factor = fitensemble(train_data, train_label, 'AdaBoostM2', 100, 'tree');
Factor = fitensemble(train_data, train_label, 'AdaBoostM2', 100, 'tree', 'type', 'classification');
Factor = fitensemble(train_data, train_label, 'Subspace', 50, 'KNN');
predict_label = predict(Factor, test_data);
[predict_label, Scores] = predict(Factor, test_data);

效果比预期差了很多。不佳。

7.判别分析分类器（discriminant analysis classifier）



Factor = ClassificationDiscriminant.fit(train_data, train_label);
Factor = ClassificationDiscriminant.fit(train_data, train_label, 'discrimType', '判别类型:伪线性...');
predict_label = predict(Factor, test_data);
[predict_label, Scores] = predict(Factor, test_data);

转自：http://blog.csdn.net/abcjennifer/article/details/17137785