Question

我之前从未使用过OpenCV，但我正在尝试编写神经网络系统来识别文本，我需要一些文本提取/分割工具。

如何使用java OpenCV预处理和分割包含文本的图像。
我不需要识别文本，我只需要将每个字母放在单独的图像中。
这样的事情：
enter image description here

Answer 1

试试这段代码。不需要OpenCV

import java.awt.image.BufferedImage;
import java.util.ArrayList;
import java.util.List;
import org.neuroph.imgrec.ImageUtilities;


public class CharExtractor {

private int cropTopY = 0;//up locked coordinate
private int cropBottomY = 0;//down locked coordinate
private int cropLeftX = 0;//left locked coordinate
private int cropRightX = 0;//right locked coordinate
private BufferedImage imageWithChars = null;
private boolean endOfImage;//end of picture
private boolean endOfRow;//end of current reading row

/**
 * Creates new char extractor with soecified text image
 * @param imageWithChars - image with text
 */
public CharExtractor(BufferedImage imageWithChars) {
    this.imageWithChars = imageWithChars;
}

public void setImageWithChars(BufferedImage imageWithChars) {
    this.imageWithChars = imageWithChars;
}

/**
 * This method scans image pixels until it finds the first black pixel (TODO: use         foreground color which is black by default).
 * When it finds black pixel, it sets cropTopY and returns true. if it reaches end of image and does not find black pixels, 
 * it sets endOfImage flag and returns false.
 * @return - returns true when black pixel is found and cropTopY value is changed, and false if cropTopY value is not changed
 */
 private boolean findCropTopY() {
    for (int y = cropBottomY; y < imageWithChars.getHeight(); y++) { // why cropYDown? -   for multiple lines of text using cropBottomY from previous line above; for first line its zero
        for (int x = cropLeftX; x < imageWithChars.getWidth(); x++) { // scan starting from the previous left crop position - or it shoud be right???
            if (imageWithChars.getRGB(x, y) == -16777216) { // if its black rixel (also consider condition close to black or not white or different from background)
                this.cropTopY = y;   // save the current y coordiante
                return true;        // and return true
            }
        }
    }
    endOfImage = true;  //sets this flag if no black pixels are found
    return false;       // and return false
}

/**
 * This method scans image pixels until it finds first row with white pixels. (TODO: background color which is white by default).
 * When it finds line whith all white pixels, it sets cropBottomY and returns true
 * @return - returns true when cropBottomY value is set, false otherwise
 */
private boolean findCropBottomY() {
    for (int y = cropTopY + 1; y < imageWithChars.getHeight(); y++) { // scan image from  top to bottom           
        int whitePixCounter = 0; //counter of white pixels in a row
        for (int x = cropLeftX; x < imageWithChars.getWidth(); x++) { // scan all pixels to right starting from left crop position
            if (imageWithChars.getRGB(x, y) == -1) {    // if its white pixel
                whitePixCounter++;                      // increase counter
            }
        }
        if (whitePixCounter == imageWithChars.getWidth()-1) { // if we have reached end of line counting white pixels (x pos)
            cropBottomY = y;// that means that we've found white line, so set current y coordinate minus 1
            return true; // as cropBottomY and finnish with true
        }
        if (y == imageWithChars.getHeight() - 1) {  // if we have reached end of image 
            cropBottomY = y;                        // set crop bottom
            endOfImage = true;                      // set corresponding endOfImage flag
            return true;                            // and return true
        }
    }
    return false;                                   // this should never happen, however its possible if image has non white bg
}

private boolean findCropLeftX() {        
    int whitePixCounter = 0;                                            // white pixel counter between the letters
    for (int x = cropRightX; x < imageWithChars.getWidth(); x++) {      // start from previous righ crop position (previous letter), and scan following pixels to the right
        for (int y = cropTopY; y <= cropBottomY; y++) {             // vertical pixel scan at current x coordinate
            if (imageWithChars.getRGB(x, y) == -16777216) {             // when we find black pixel
                cropLeftX = x;                                          // set cropLeftX
                return true;                                            // and return true
            }
        }

        // BUG?: this condition looks strange.... we might not need whitePixCounter at all, it might be used for 'I' letter
        whitePixCounter++;                                              // if its not black pixel assume that its white pixel
        if (whitePixCounter == 3) {                                     // why 3 pixels? its hard coded for some case and does not work in general...!!!
            whitePixCounter = 0;                                        // why does it sets to zero, this has no purporse at all...
        }
    }
    endOfRow = true;        // if we have reached end of row and we have not found black pixels, set the endOfRow flag
    return false;           // and return false
}

/**
 * This method scans image pixels to the right until it finds next row where all pixel are white, y1 and y2.
 * @return - return true  when x2 value is changed and false when x2 value is not changed
 */
private boolean findCropRightX() {
    for (int x = cropLeftX + 1; x < imageWithChars.getWidth(); x++) {   // start from current cropLeftX position and scan pixels to the right
        int whitePixCounter = 0;
        for (int y = cropTopY; y <= cropBottomY; y++) {             // vertical pixel scan at current x coordinate
            if (imageWithChars.getRGB(x, y) == -1) {                    // if we have white pixel at current (x, y)
                whitePixCounter++;                                      // increase whitePixCounter
            }
        }

        // this is for space!
        int heightPixels = cropBottomY - cropTopY;                      // calculate crop height
        if (whitePixCounter == heightPixels+1) {                         // if white pixel count is equal to crop height+1  then this is white vertical line, means end of current char/ (+1 is for case when there is only 1 pixel; a 'W' bug fix)
            cropRightX = x;                                             // so set cropRightX    
            return true;                                                // and return true
        }

        // why we need this when we allready have condiiton in the for loop? - for the last letter in the row.
        if (x == imageWithChars.getWidth() - 1) {                       // if we have reached end of row with x position    
            cropRightX = x;                                             // set cropRightX
            endOfRow = true;                                            // set endOfRow flag
            return true;                                                // and return true
        }
    }       
}

public List<BufferedImage> extractCharImagesToRecognize() {
    List<BufferedImage> trimedImages = new ArrayList<BufferedImage>();
    int i = 0;

    while (endOfImage == false) {
        endOfRow = false;
        boolean foundTop = findCropTopY();
        boolean foundBottom = false;
        if (foundTop == true) {
            foundBottom = findCropBottomY();
            if (foundBottom == true) {
                while (endOfRow == false) {
                    boolean foundLeft = false;
                    boolean foundRight = false;
                    foundLeft = findCropLeftX();
                    if (foundLeft == true) {
                        foundRight = findCropRightX();
                        if (foundRight == true) {
                            BufferedImage image = ImageUtilities.trimImage(ImageUtilities.cropImage(imageWithChars, cropLeftX, cropTopY, cropRightX, cropBottomY));
                            trimedImages.add(image);                                
                            i++;                                                                               
                        }
                    }
                }
                cropLeftX = 0;
                cropRightX = 0;
            }
        }
    }
    cropTopY = 0;
    cropBottomY = 0;
    endOfImage = false;

    return trimedImages;
}


public static void main(String[] args) throws Exception {
    File f=new File("./written.png");
    BufferedImage img=ImageIO.read(f);
    CharExtractor ch=new CharExtractor(img);
    List<BufferedImage> list=ch.extractCharImagesToRecognize();

    for(int i=0;i<list.size();i++)
    {
         File outputfile = new File("./char_" +i+ ".png");
         ImageIO.write(list.get(i),"png", outputfile);
    }        
}
}

Answer 2

您要做的是一般场景文本本地化问题，而且非常困难。看看这篇文章的灵感 - http://www.maseltov.eu/wp-content/uploads/2014/02/CTU-03_Real-Time-Scene-Text-Localization-and-Recognition.pdf

你能做的是：

编写一个从图像中提取MSER对象的程序
从每个MSER确定的每个补丁中提取特征（文章中概述了哪些特征）
训练你的分类器（在你的情况下是一个神经网络，我猜？），以便它能够区分角色和非角色区域
编写一个程序，使用您的分类器提取MSER并使用经过培训的NN对它们进行分类。

MSER算法在OpenCV中实现，因此这是一个优点。那里也有神经网络分类器，但由于我只使用SVM，所以我不能过多评论。我应该说我们必须解决这个问题，使用OpenCV是完全可能的。只是不希望把所有东西放在银盘上 - 涉及很多工作;特别是在选择和提取blob特征时。

Answer 3

我不熟悉神经网络，但如果您只是想在图像中找到关于比例和旋转的字母，我可以推荐这个项目http://www.codeproject.com/Articles/196168/Contour-Analysis-for-Image-Recognition-in-C 它是用C＃编写的，但您可以将其移植到java或至少获得有关此主题的大量见解。

文本提取和分段打开CV

3 个答案: