Java Neural Network Framework Neuroph / Discussion / Help: Strange results of image detection

Hello,

first of all, thank you very much for developing such great environment. I'm trying to create a neural network to detect images. Here is detail info what I have installed:
Product Version: NeurophStudio 201408251540
Java: 1.7.0_71; Java HotSpot(TM) Server VM 24.71-b01
Runtime: Java(TM) SE Runtime Environment 1.7.0_71-b14
System: Windows 7 version 6.1 running on x86; Cp1250; sk_SK (neurophstudio)

Following the tutorial at http://neuroph.sourceforge.net/image_recognition.html I was able to create a network, learn it and detect a picture (in NeurophStudio). Then I was able to load created network in java and detect picture. Now I'm trying to create a network runtime in java, learn it and detect a picture. But the results are not very good.

I created an example with 3 neural networks:
1. network - java runtime created network
2. network - created in NeurophStudio, but without learned data
3. network - created in NeurophStudio with learned data

Here is the source code:


import org.apache.commons.io.FilenameUtils;

import org.neuroph.core.NeuralNetwork;

import org.neuroph.core.data.DataSet;

import org.neuroph.imgrec.ColorMode;

import org.neuroph.imgrec.FractionRgbData;

import org.neuroph.imgrec.ImageRecognitionHelper;

import org.neuroph.imgrec.ImageRecognitionPlugin;
import org.neuroph.imgrec.ImageUtilities;

import org.neuroph.imgrec.image.Dimension;

import org.neuroph.nnet.learning.MomentumBackpropagation;

import org.neuroph.util.TransferFunctionType;
import java.text.DecimalFormat;

import java.text.NumberFormat;

import java.util.ArrayList;

import java.util.HashMap;

import java.util.List;

import java.util.Map;

import java.awt.image.BufferedImage;

import java.io.File;

import java.io.IOException;
public class main {
public static void main(String[] args) {

    //
    // Create learning data
    //
    File folder = new File ("Pictures\\");
    List<String> imageLabels = new ArrayList<String> ();

    HashMap<String, BufferedImage> imagesMap = new HashMap<String, BufferedImage> ();
    for (File file : folder.listFiles ())
    {
        imageLabels.add(FilenameUtils.removeExtension(file.getName()));
        imagesMap.put(file.getName(), ImageUtilities.resizeImage (ImageUtilities.loadImage(file), 20, 20));           
    }
    Map<String, FractionRgbData> imageRgbData = ImageUtilities.getFractionRgbDataForImages (imagesMap);
    DataSet learningData = ImageRecognitionHelper.createTrainingSet(imageLabels, imageRgbData);

    // 
    // create neural networks
    //  
    ArrayList<Integer> layers = new ArrayList<Integer> ();layers.add (12);
    NeuralNetwork nn1 = ImageRecognitionHelper.createNewNeuralNetwork ("recognition", new Dimension (20, 20), ColorMode.FULL_COLOR, imageLabels, layers, TransferFunctionType.SIGMOID); // create my own network
    NeuralNetwork nn2 = NeuralNetwork.createFromFile("NeuralNetworks\\test_notTrained.nnet");   // load network created in NeurophStudio following tutorial at http://neuroph.sourceforge.net/image_recognition.html
    NeuralNetwork nn3 = NeuralNetwork.createFromFile("NeuralNetworks\\test_trained.nnet");      // load network created in NeurophStudio following tutorial at http://neuroph.sourceforge.net/image_recognition.html

    //
    // learn data
    //
    MomentumBackpropagation mb1 = (MomentumBackpropagation)nn1.getLearningRule();
    mb1.setLearningRate(0.2);
    mb1.setMaxError(0.1);
    mb1.setMomentum(0.7);

    MomentumBackpropagation mb2 = (MomentumBackpropagation)nn2.getLearningRule();
    mb2.setLearningRate(0.2);
    mb2.setMaxError(0.1);
    mb2.setMomentum(0.7);

    nn1.learn(learningData);
    nn2.learn(learningData);

    // get the image recognition plugin from neural network    
    ImageRecognitionPlugin ir1 = (ImageRecognitionPlugin)nn1.getPlugin(ImageRecognitionPlugin.class); 
    ImageRecognitionPlugin ir2 = (ImageRecognitionPlugin)nn2.getPlugin(ImageRecognitionPlugin.class);
    ImageRecognitionPlugin ir3 = (ImageRecognitionPlugin)nn3.getPlugin(ImageRecognitionPlugin.class);

    //
    // Try to check all learning images
    //
    System.out.println("-----------------------------------------------------------------------------------------------------");
    for (File file : folder.listFiles ())
    {
        System.out.println("Checking '" + FilenameUtils.removeExtension(file.getName()) + "'");
        System.out.println("Java Runtime created network     : " + RecognizeImage(ir1, file.getPath()));
        System.out.println("NS created network (not trained) : " + RecognizeImage(ir2, file.getPath()));
        System.out.println("NS created network (trained)     : " + RecognizeImage(ir3, file.getPath()));
        System.out.println("-----------------------------------------------------------------------------------------------------");
    }
}

/**
 * 
 * @param imageRecognition
 * @param imagePath
 */
private static String RecognizeImage(ImageRecognitionPlugin imageRecognition, String imagePath) {
    String result = "";
    try {
        HashMap<String, Double> output = imageRecognition.recognizeImage(new File(imagePath));

        NumberFormat formatter = new DecimalFormat("#0.0"); 
        double maxPercent = Double.MIN_VALUE;
        for (Map.Entry<String, Double> entry : output.entrySet()) {

            if ( entry.getValue() > maxPercent){
                maxPercent = entry.getValue(); 
                result = entry.getKey() + " (" + formatter.format(maxPercent * 100) + " %)";
            }
        }

    } catch(IOException ioe) {
     ioe.printStackTrace();
    }

    return result;
}

}

And here are the results:
Neuron layer size counts vector = [1200, 12, 4] apr 30, 2015 12:21:01 PM org.neuroph.core.learning.LearningRule onStart INFO: Learning Started apr 30, 2015 12:21:01 PM org.neuroph.core.learning.LearningRule onStop INFO: Learning Stoped apr 30, 2015 12:21:01 PM org.neuroph.core.learning.LearningRule onStart INFO: Learning Started apr 30, 2015 12:21:01 PM org.neuroph.core.learning.LearningRule onStop INFO: Learning Stoped



Checking 'bird'

Java Runtime created network : bird (64,8 %)

NS created network (not trained) : bird (45,0 %)

NS created network (trained) : bird (89,0 %)

Checking 'cat'

Java Runtime created network : bird (64,7 %)

NS created network (not trained) : bird (58,0 %)

NS created network (trained) : cat (92,7 %)

Checking 'dog'

Java Runtime created network : dog (70,3 %)

NS created network (not trained) : dog (73,7 %)

NS created network (trained) : dog (88,2 %)

Checking 'rose'

Java Runtime created network : rose (60,8 %)

NS created network (not trained) : rose (47,7 %)

NS created network (trained) : rose (91,5 %)

Besides the numbers are differemt, the 'cat' picture is wrongly detected. I suspect I did something wrong in the learning process, but I cannot figure it out.

Could you please advice what can be wrong?

Last edit: Andrej K 2015-04-30