OpenCV for Java KNN (KNearest-K nearest neighbor algorithm) example code



OpenCV + Java KNN ( KNearest-K nearest neighbor algorithm) example code

This article gives an example of the OpenCV machine learning K -nearest neighbor algorithm implemented in the Java language.

OpenCV version: 3.4.0 . There are no tutorials or examples about Java language ML in the official documents . There are only code examples of C++ and Python . I found an example on the Internet, passed the actual test, and attached Chinese notes.

Original address: https://stackoverflow.com/questions/32612058/java-opencv-find-k-nearest-neighbor-c-to-java-conversion

Sample code:

import java.text.DecimalFormat; import java.util.*;   import org.opencv.core.*; import org.opencv.imgcodecs.Imgcodecs; import org.opencv.ml.KNearest; import org.opencv.ml.Ml; import org.opencv.utils.Converters;   public class Main {      public static void main(String[] args) {         // TODO Auto-generated method stub         // load the OpenCV native library         System.loadLibrary(Core.NATIVE_LIBRARY_NAME);           // The path of the test data set after the OpenCV installation package is decompressed: samples/data/digits.png         // Use the digital image dataset attached to the OpenCV installation package. The dataset has a total of 5000 digital samples / samples from 0 to 9 , and the image resolution is 2000x1000 .         // The size of each digital sample is 20x20 , so there are 2000/20=100 digital samples in each line, each number from 0 to 9 occupies 5 lines in turn, and 10 numbers have a total of 50 lines.         // Pay attention to the image path, you can use an absolute path         Mat digits = Imgcodecs.imread("E:\\...\\images\\digits.png", 0);                 // setup train/test data:         Mat trainData = new Mat();         Mat testData = new Mat();         List<Integer> trainLabs = new ArrayList<Integer>();         List<Integer> testLabs = new ArrayList<Integer>();           // 10 digits a 5 rows:         // 一次行读入所有数据，训练数据和测试数据各一半         // 数据集共 50 行         for (int r=0; r<50; r++) {             // 100 digits per row:             // 每行 100 个数字样例             for (int c=0; c<100; c++) {             // crop out 1 digit:             // 每次读入一个数字样本，大小：20x20             Mat num = digits.submat(new Rect(c*20,r*20,20,20));             // we need float data for knn:             // knn算法的输入为浮点型，在此转换             num.convertTo(num, CvType.CV_32F);             // 50/50 train/test split:             if (c % 2 == 0) {                  // 偶数行作为训练样本                  // for opencv ml, each feature has to be a single row:                  // OpenCV中的ml算法要求输入训练数据的每一个样本（此例中可理解为样本特征：feature）占据一行                  // num 本身为 20x20 的矩阵，转换为 1 行 n列的矩阵，其中 n=20x20=400                  trainData.push_back(num.reshape(1,1));                  // add a label for that feature (the digit number):                  // 对应每个训练样本建立对应的Label（正确答案）                  trainLabs.add(r/5);             } else {                  // 奇数行作为测试数据                  testData.push_back(num.reshape(1,1));                  testLabs.add(r/5);             }             }                        }           // make a Mat of the train labels, and train knn:         // 构建 KNearest 对象，3.0版本之前似乎可以直接 new 一个         KNearest knn = KNearest.create();         // train函数原型定义为：public  boolean train(Mat samples, int layout, Mat responses)         // 其中第一个参数自不必说，是输入的训练数据，其中每一行代表一个样本（特征）；第二个参数是指定样本的布局，是每行一个样本，         // 还是每列一个样本，没有默认值；第三个参数为训练样本对应的正确答案（label）         // 扩展：Converters类中有OpenCV提供的许多数据（类型）转换工具，非常实用         knn.train(trainData, Ml.ROW_SAMPLE, Converters.vector_int_to_Mat(trainLabs));                 // 使用测试数据集进行测试         // 测试数据也是每行一个样本         // now test predictions:         int err = 0;         for (int i=0; i<testData.rows(); i++)         {             // 读取一行（一个测试样本）             Mat one_feature = testData.row(i);             // 预期的（标准/正确）答案             int testLabel = testLabs.get(i);               Mat res = new Mat();             // 查找匹配：第一个参数为输入样本（可一次输入多个样本），第二个参数为需要返回的K个邻近（即KNearest的那个K），第三个参数为返回             // 结果（res: result），结果为样本对应的Label值，每一个样本的匹配结果对应一行。             // 如果输入仅有一个样本，则返回结果（p）就是预测结果。参数 1即为K-近邻算法的关键参数 K！             float p = knn.findNearest(one_feature, 1, res);             System.out.println(testLabel + " " + p + " " + res.dump());                         // 统计识别误差             int iRes = (int) p;             if(iRes != testLabel) {                 err++;             }         }           // 输出（屏幕提示）识别精度         float accuracy = (float) ((2500 - (float)err) / 2500.0);         DecimalFormat df = new DecimalFormat("0.0000");         System.out.println("error count: " + err + ", accuracy is: " + df.format(accuracy));           // 以下为源作者一些说明，大意是如果在实际应用中需要借助OpenCV自带的数字样本集训练的模型识别数字，需要做的一些工作         // 在此不作具体解释，有疑问的可以留言         hmm, the 'real world' test case probably looks more like this:         make sure, you follow the very same pre-processing steps used in the train phase:         //  Mat one_feature = Imgcodecs.imread("one_digit.png", 0);         //  Mat feature;         //  one_feature.convertTo(feature, CvTypes.CV_32F);         //  Imgproc.resize(feature, feature, Size(20,20));         //  int predicted = knn.findNearest(feature.reshape(1,1), 1);       }   }

代码在本人主机上的输出结果为： error count: 154, accuracy is: 0.9384

可见识别正确率仅有约 93.8%