Java基于OpenCV开发ANN-预测手写数字的案例

本博客属于原创，若转载请标明转载出处：https://blog.csdn.net/qq_44091004/article/details/90573495
小编在开发的过程中就省去了ANN的原理部分，所以小编认为大家是有一定基础的。本案例是ANN-预测手写数字的案例。
本文需要的开发环境是新建一个工程（new project），这个工程里已经配置了OpenCV开发环境。
第一步：首先就是数据集材料的从哪里来，小编这里有一段程序是可以直接运行的App,作用是通过一个面板手绘数字，然后通过缩小的功能将图片都缩小为12*12的尺寸，然后存在自己指定的盘中，比如小编这里将手写图片数据都存在E://opencv3.1//samples//ocr中。从下图中可以看到，我的源数据集从何而来。

第二步：建立手写数字的资料库。
该数据库里包含五十个训练数据，每个数字包含五个手写体，还有一个测试数据sample7。下面的代码中有一些需要注意的地方，首先是将图片这个二维矩阵数据转换为行样本存储在Mat中，ANN中每个样本的标签用一个1行10列的矩阵来存储。

package com;

import org.opencv.core.Core;
import org.opencv.core.CvType;
import org.opencv.core.Mat;
import org.opencv.imgcodecs.Imgcodecs;

public class OcrDatabase {
	static {System.loadLibrary(Core.NATIVE_LIBRARY_NAME);};
	Mat trainingLabelsFloatMat = new Mat(50,10,CvType.CV_32FC1);//创建训练数据集的标签存储矩阵
	Mat trianingDataMat = new Mat(50,144,CvType.CV_32FC1) ;//创建训练数据集的存储矩阵
	Mat testingDataMat = new Mat(10,144,CvType.CV_32FC1);//创建测试数据集的存储矩阵

	float[] testingLabels = {0,1,2,3,4,5,6,7,8,9};
	float[][] trainingLabelsFloat = {
			{1,0,0,0,0,0,0,0,0,0},
			{1,0,0,0,0,0,0,0,0,0},
			{1,0,0,0,0,0,0,0,0,0},
			{1,0,0,0,0,0,0,0,0,0},
			{1,0,0,0,0,0,0,0,0,0},
			{0,1,0,0,0,0,0,0,0,0},
			{0,1,0,0,0,0,0,0,0,0},
			{0,1,0,0,0,0,0,0,0,0},
			{0,1,0,0,0,0,0,0,0,0},
			{0,1,0,0,0,0,0,0,0,0},
			{0,0,1,0,0,0,0,0,0,0},
			{0,0,1,0,0,0,0,0,0,0},
			{0,0,1,0,0,0,0,0,0,0},
			{0,0,1,0,0,0,0,0,0,0},
			{0,0,1,0,0,0,0,0,0,0},
			{0,0,0,1,0,0,0,0,0,0},
			{0,0,0,1,0,0,0,0,0,0},
			{0,0,0,1,0,0,0,0,0,0},
			{0,0,0,1,0,0,0,0,0,0},
			{0,0,0,1,0,0,0,0,0,0},
			{0,0,0,0,1,0,0,0,0,0},
			{0,0,0,0,1,0,0,0,0,0},
			{0,0,0,0,1,0,0,0,0,0},
			{0,0,0,0,1,0,0,0,0,0},
			{0,0,0,0,1,0,0,0,0,0},
			{0,0,0,0,0,1,0,0,0,0},
			{0,0,0,0,0,1,0,0,0,0},
			{0,0,0,0,0,1,0,0,0,0},
			{0,0,0,0,0,1,0,0,0,0},
			{0,0,0,0,0,1,0,0,0,0},
			{0,0,0,0,0,0,1,0,0,0},
			{0,0,0,0,0,0,1,0,0,0},
			{0,0,0,0,0,0,1,0,0,0},
			{0,0,0,0,0,0,1,0,0,0},
			{0,0,0,0,0,0,1,0,0,0},
			{0,0,0,0,0,0,0,1,0,0},
			{0,0,0,0,0,0,0,1,0,0},
			{0,0,0,0,0,0,0,1,0,0},
			{0,0,0,0,0,0,0,1,0,0},
			{0,0,0,0,0,0,0,1,0,0},
			{0,0,0,0,0,0,0,0,1,0},
			{0,0,0,0,0,0,0,0,1,0},
			{0,0,0,0,0,0,0,0,1,0},
			{0,0,0,0,0,0,0,0,1,0},
			{0,0,0,0,0,0,0,0,1,0},
			{0,0,0,0,0,0,0,0,0,1},
			{0,0,0,0,0,0,0,0,0,1},
			{0,0,0,0,0,0,0,0,0,1},
			{0,0,0,0,0,0,0,0,0,1},
			{0,0,0,0,0,0,0,0,0,1}
	};

	Mat sample7=new Mat(1,144,CvType.CV_32FC1);
	public OcrDatabase() {
		Mat source;
		//assign training Mat
		for(int i = 0;i < 50;i++) {
			if(i < 10) {
				source = Imgcodecs.imread("E://opencv3.1//samples//ocr//0"+i+".jpg",Imgcodecs.CV_LOAD_IMAGE_GRAYSCALE);
			}
			else{
				source = Imgcodecs.imread("E://opencv3.1//samples//ocr//"+i+".jpg",Imgcodecs.CV_LOAD_IMAGE_GRAYSCALE);
			}
			Mat temp = source.reshape(1, 144);
			for(int j = 0;j < 144;j++) {
				double[] data = new double[1];
				data = temp.get(j, 0);
				trianingDataMat.put(i, j, data);
			}
			trainingLabelsFloatMat.put(i, 0, trainingLabelsFloat[i]);
		}

		Mat sample = Imgcodecs.imread("E://opencv3.1//samples//ocr//number71.jpg",Imgcodecs.CV_LOAD_IMAGE_GRAYSCALE);
		Mat tempSample7=sample.reshape(1,144);
		for(int j=0;j<144;j++){
			double[] data=new double[1];
			data=tempSample7.get(j, 0);
			sample7.put(0, j, data);
		}
	}

	public Mat getTrainingDataMat() {
		return trianingDataMat;
	}

	public void setTrainingDataMat(Mat trainingDataMat) {
		this.trianingDataMat = trainingDataMat;
	}

	public Mat getTrainingLabelsFloatMat() {
		return trainingLabelsFloatMat;
	}

	public void setTrainingLabelsFloatMat(Mat trainingLabelsFloatMat) {
		this.trainingLabelsFloatMat = trainingLabelsFloatMat;
	}


}

第三步：测试类

import org.opencv.core.Core;
import org.opencv.ml.ANN_MLP;
import org.opencv.ml.Ml;
import org.opencv.core.CvType;
import org.opencv.core.Mat;
import org.opencv.core.TermCriteria;

public class KNN {
	static{ System.loadLibrary(Core.NATIVE_LIBRARY_NAME); }
	public static void main(String[] args) {
		OcrDatabase ocr=new OcrDatabase();//创建数据集对象，相当于做试验前准备好的实验数据，封装了起来，以便调用
		ANN_MLP ann=ANN_MLP.create();//创建ANN
		//设置网络的模型，首先创建网络模型对象，然后创建数组，将数组中的数据存储到Mat对象中，分别包含输入层，隐含层和输出层，其中输入
		//层为144个特征，隐含层包括两层，分别有20和10个神经元，输出层为10个数据
		Mat layerSize=new Mat(4,1,CvType.CV_32SC1);

		int[] layerSizeAry={144, 20,10, 10};

		layerSize.put(0,0,layerSizeAry[0]);
		layerSize.put(1,0,layerSizeAry[1]);
		layerSize.put(2,0,layerSizeAry[2]);
		layerSize.put(3,0,layerSizeAry[3]);
		ann.setLayerSizes(layerSize);

		ann.setTrainMethod(ann.BACKPROP);//设置训练方法：误差反向传播
		TermCriteria criteria=new TermCriteria(TermCriteria.MAX_ITER|TermCriteria.EPS, 300, 0.001);//创建对象，设置标准属性包括最大迭代次数和误差率
		ann.setTermCriteria(criteria);
		ann.setActivationFunction(ann.SIGMOID_SYM);//设置激励函数为SIGMOLD类型

		boolean r=ann.train(ocr.getTrainingDataMat(), Ml.ROW_SAMPLE, ocr.getTrainingLabelsFloatMat());//训练函数，样本类型为行样本，然后获取训练数据集和标签进行训练
		System.out.println("是否有训练成功="+r);

		//测试sample7
		float result7= ann.predict(ocr.sample7);
		System.out.println("预测7结果="+result7);
	}

}

运行结果如下图所示：

最后，作者参考的书籍：
《opencv3 使用java开发手册》