ビール画像判定プログラムできたよ！

画像同士の距離が取れるなら、ビール画像の判定プログラムもできるよね。
ってことで、手持ちの画像をSVMに食わせてみて、ビール画像を判定させてみました。

誤判別？なにそれ

ソースはこんな感じ。
ビールの画像はbeerがついたフォルダかファイル名にしてください。

import java.awt.*;
import java.awt.image.BufferedImage;
import java.io.*;
import java.util.*;
import java.util.List;
import javax.imageio.ImageIO;
import javax.swing.ImageIcon;
import javax.swing.text.*;

public class ColorCoherenceVecotrLaern extends javax.swing.JFrame {

    /** Creates new form ColorCoherenceVecotrLaern */
    public ColorCoherenceVecotrLaern() {
        initComponents();
        System.out.println("CCV計算");
        readFile(new File("C:\\search"));
        System.out.println("カーネルパラメータ");
        solveGaussianParam();
        System.out.println("SMO");
        learn();
        System.out.println("終了");
    }

    @SuppressWarnings("unchecked")
    private void initComponents() {

        javax.swing.JScrollPane jScrollPane1 = new javax.swing.JScrollPane();
        tpResut = new javax.swing.JTextPane();
        javax.swing.JPanel jPanel1 = new javax.swing.JPanel();
        txtSearch = new javax.swing.JTextField();
        javax.swing.JButton btnJudge = new javax.swing.JButton();

        setDefaultCloseOperation(javax.swing.WindowConstants.EXIT_ON_CLOSE);

        jScrollPane1.setViewportView(tpResut);

        getContentPane().add(jScrollPane1, java.awt.BorderLayout.CENTER);

        txtSearch.setColumns(30);
        jPanel1.add(txtSearch);

        btnJudge.setText("判定");
        btnJudge.addActionListener(new java.awt.event.ActionListener() {
            public void actionPerformed(java.awt.event.ActionEvent evt) {
                btnJudgeActionPerformed(evt);
            }
        });
        jPanel1.add(btnJudge);

        getContentPane().add(jPanel1, java.awt.BorderLayout.NORTH);

        pack();
    }

    private void btnJudgeActionPerformed(java.awt.event.ActionEvent evt) {
        String filename = txtSearch.getText();
        File f = new File(filename);
        try {
            BufferedImage img = ImageIO.read(f);
            int[] ccv = colorCoherenceVector(img);
            int result = trial(ccv);
            int wd = img.getWidth();
            int ht = img.getHeight();
            if(wd > ht){
                ht = 80 * ht / wd;
                wd = 80;
            }else{
                wd = 80 * wd / ht;
                ht = 80;
            }
            BufferedImage thumb = new BufferedImage(wd, ht, BufferedImage.TYPE_INT_RGB);
            Graphics2D g = (Graphics2D) thumb.getGraphics();
            g.setRenderingHint(RenderingHints.KEY_ANTIALIASING, RenderingHints.VALUE_ANTIALIAS_ON);
            g.drawImage(img, 0, 0, wd, ht, this);
            g.dispose();
            StyledDocument sd = tpResut.getStyledDocument();
            SimpleAttributeSet sas = new SimpleAttributeSet();
            StyleConstants.setIcon(sas, new ImageIcon(thumb));
            sd.insertString(sd.getLength(), "画像", sas);
            sd.insertString(sd.getLength(), result > 0 ? "ビールだ！\n" : "なにそれ\n", null);
        }catch(BadLocationException e){
        }catch(IOException e){
            System.out.println(e.getMessage());
        }

    }

    //ファイルのCCVを求める
    private void readFile(File f){
        if(f.isDirectory()){
            File[] files = f.listFiles();
            for(File file : files){
                readFile(file);
            }
        }else{
            try {
                BufferedImage img = ImageIO.read(f);
                if(img == null) return;
                int[] data = colorCoherenceVector(img);
                int i = -1;
                if(f.getCanonicalPath().contains("beer")){
                    i = 1;
                }
                patterns.add(new AbstractMap.SimpleEntry<Integer, int[]>(i, data));
            } catch (IOException ex) {
                System.out.println(ex.getMessage());
            }

        }
    }

    /** CCVを求める */
    public static int[] colorCoherenceVector(BufferedImage imgsrc){

        int w = imgsrc.getWidth();
        int h = imgsrc.getHeight();
        //サイズ正規化
        int limit = 200;
        if(w < h){
            w = w * limit / h;
            h = limit;
        }else{
            h = h * limit / w;
            w = limit;
        }
        BufferedImage img = new BufferedImage(w, h, BufferedImage.TYPE_INT_RGB);
        Graphics2D grp = (Graphics2D) img.getGraphics();
        grp.setRenderingHint(RenderingHints.KEY_ANTIALIASING, RenderingHints.VALUE_ANTIALIAS_ON);
        grp.drawImage(imgsrc, 0, 0, w, h, null);
        grp.dispose();

        //ガウシアンフィルタ
        int[] ctemp = img.getRGB(0, 0, w, h, null, 0, w);
        int[] ctbl = new int[ctemp.length];
        int[][] filter = {
            {1, 2, 1},
            {2, 4, 2},
            {1, 2, 1}};
        for(int y = 0; y < h; ++y){
            for(int x = 0; x < w; ++x){
                int tr = 0;
                int tg = 0;
                int tb = 0;
                int t = 0;
                for(int i = -1; i < 2; ++i){
                    for(int j = -1; j < 2; ++j){
                        if(y + i < 0) continue;
                        if(x + j < 0) continue;
                        if(y + i >= h) continue;
                        if(x + j >= w) continue;
                        t += filter[i + 1][j + 1];
                        int adr = (x + j) + (y + i) * w;
                        tr += filter[i + 1][j + 1] * ((ctemp[adr] >> 16) & 255);
                        tg += filter[i + 1][j + 1] * ((ctemp[adr] >> 8)  & 255);
                        tb += filter[i + 1][j + 1] * ( ctemp[adr]        & 255);
                    }
                }
                ctbl[x + y * w] = ((tr / t) << 16) + ((tg / t) << 8) + tb / t;
            }
        }

        //減色
        for(int i = 0; i < ctbl.length; ++i){
            int r = (ctemp[i] >> 16) & 192;
            int g = (ctemp[i] >> 8) & 192;
            int b = ctemp[i] & 192;
            ctbl[i] = (r << 16) + (g << 8) + b;
        }

        //タグ付け
        int[][] lbl = new int[w][h];
        int id = 0;
        for(int y = 0; y < h; ++y){
            for(int x = 0; x < w; ++x){
                int col = ctbl[y * w + x];
                if(y > 0){
                    if(x > 0){
                        if(ctbl[(y - 1) * w + x - 1] == col){
                            //左上と一緒
                            lbl[x][y] = lbl[x - 1][y - 1];
                            continue;
                        }
                    }
                    if(ctbl[(y - 1) * w + x] == col){
                        //上と一緒
                        lbl[x][y] = lbl[x][y - 1];
                        continue;
                    }
                    if(x < w - 1){
                        if(ctbl[(y - 1) * w + x + 1] == col){
                            //右上と一緒
                            lbl[x][y] = lbl[x + 1][y - 1];
                            continue;
                        }
                    }
                }
                if(x > 0){
                    if(ctbl[y * w + x - 1] == col){
                        //左と一緒
                        lbl[x][y] = lbl[x - 1][y];
                        continue;
                    }
                }
                lbl[x][y] = id;
                ++id;
            }
        }
        //集計
        int[] count = new int[id];
        int[] color = new int[id];
        for(int x = 0; x < w; ++x){
            for(int y = 0; y < h; ++y){
                count[lbl[x][y]]++;
                color[lbl[x][y]] = ctbl[y * w + x];
            }
        }
        int[] data = new int[129];
        for(int i = 0; i < id; ++i){
            int d = color[i];
            color[i] = (((d >> 22) & 3) << 4) + (((d >> 14) & 3) << 2) + ((d >> 6) & 3);
            if(count[i] < 20){
                data[color[i] * 2 + 1] ++;
            }else{
                data[color[i] * 2] ++;
            }
        }
        return data;
    }


    double sig = 35;
    /** カーネルパラメータを求める */
    void solveGaussianParam(){
        double s = 0;
        int count = 0;
        for(Map.Entry<Integer, int[]> p1 : patterns){
            double m = Double.MAX_VALUE;
            for(Map.Entry<Integer, int[]> p2 : patterns){
                double d = 0;
                for(int i = 0; i < p1.getValue().length; ++i){
                    int t = p1.getValue()[i] - p2.getValue()[i];
                    d += t * t;
                }
                if(d == 0) continue;
                if(d < m) m = d;
            }
            if(m == Double.MAX_VALUE) continue;
            s += m;
            ++count;
        }
        sig = Math.sqrt(s / count) * 2;
        System.out.println(sig);
    }

    double kernel(int[] x1, int[] x2){
        //ガウシアンカーネル
        int total = 0;
        for(int i = 0; i < x1.length; ++i){
            total += (x1[i] - x2[i]) * (x1[i] - x2[i]);
        }
        return Math.exp(-total / (sig * sig));
    }

    double[] w;//係数
    double b;//バイアス
    final double c = 10;//許容範囲？無限大にするとハードマージンになるはずだけど
    final double tol = 0.7;//KKT条件の許容範囲(1 - ε)
    double[] lambda;
    double z = 0;
    List<Map.Entry<Integer, int[]>> patterns =
            new ArrayList<Map.Entry<Integer, int[]>>();
    public void learn() {
        w = new double[patterns.size()];
        b = 0;

        lambda = new double[patterns.size()];
        for(int i = 0; i < lambda.length; ++i){
            lambda[i] = 0;
        }

        //未定乗数を求める
        boolean alldata = true;//すべてのデータを処理する場合
        boolean changed = false;//変更があった
        eCache = new double[patterns.size()];
        int lp;
        for(lp = 0; lp < 500000 && (alldata || changed); ++lp)  {
            changed = false;
            z = 0;
            boolean lastchange = true;
            PROC_LOOP:
            for(int j = 0; j < patterns.size(); ++j){
                //基準点2を選ぶ
                double alpha2 = lambda[j];
                if(!alldata && (alpha2 <= 0 || alpha2 >= c)){// 0 < α < C の点だけ処理する
                    continue;
                }
                if(lastchange){
                    //初回やデータがかわったときキャッシュのクリア
                    for(int i = 0; i < eCache.length; ++i) eCache[i] = Double.NaN;
                }
                lastchange = false;

                int t2 = patterns.get(j).getKey();
                double fx2 = calcE(j);

                //KKT条件の判定
                double r2 = fx2 * t2;
                if(!((alpha2 < c && r2 < -tol) || (alpha2 > 0 && r2 > tol))){//KKT条件をみたすなら処理しない
                    continue;
                }
                //基準点1を選ぶ
                //選択法1
                int i = 0;
                int offset = (int)(Math.random() * patterns.size());

                double max = -1;
                for(int ll = 0; ll < patterns.size(); ++ll){//全データにつき
                    int l = (ll + offset) % patterns.size();
                    //0 < α < C
                    if(0 >= lambda[l] || c <= lambda[l]) continue;
                    double dif = Math.abs(calcE(l) - fx2);
                    if(dif > max){
                        max = dif;
                        i = l;
                    }
                }
                if(max >= 0){
                    if(step(i, j)){
                        //処理をしたら次へ
                        changed = true;
                        lastchange = true;
                        continue PROC_LOOP;
                    }
                }
                //選択法2
                offset = (int)(Math.random() * patterns.size());//ランダムな位置から
                for(int l = 0; l < patterns.size(); ++l){
                    //0 < α < C
                    i = (l + offset) % patterns.size();
                    if(0 >= lambda[i] || c <= lambda[i]) continue;
                    if(step(i, j)){
                        //処理をしたら次へ
                        changed = true;
                        lastchange = true;
                        continue PROC_LOOP;
                    }
                }
                //選択法3
                offset = (int)(Math.random() * patterns.size());//ランダムな位置から
                for(int l = 0; l < patterns.size(); ++l){
                    i = (l + offset) % patterns.size();
                    if(step(i, j)){
                        //処理をしたら次へ
                        changed = true;
                        lastchange = true;
                        continue PROC_LOOP;
                    }
                }
            }

            ////すべてのデータを処理しても処理するものがなければ終了になる
            if(z < 0.01) changed = false;
            if(alldata){
                alldata = false;
            }else{
                if(changed) alldata = true;
            }
        }
        //System.out.println(lp);

        //wの値を求める
        for(int i = 0; i < w.length; ++i){
            w[i] = lambda[i] * patterns.get(i).getKey();
        }
        //bを求める
        b = 0;
        int count = 0;
        for(int i = 0; i < lambda.length; ++i){
            if(Math.abs(w[i]) <= 0.05) continue;
            for(int l = 0; l < w.length; ++l){
                b -= w[l] * kernel(
                        patterns.get(i).getValue(), patterns.get(l).getValue());
            }
            ++count;
        }
        b /= count;
    }

    public int trial(int[] data) {
        double s = b;
        for(int i = 0; i < w.length; ++i){
            Map.Entry<Integer, int[]> p = patterns.get(i);
            s += w[i] * kernel(data, p.getValue());
        }
        return s > 0 ? 1 : -1;
    }

    private double[] eCache;
    private double calcE(int i){
        if(!Double.isNaN(eCache[i])) return eCache[i];
        double e = b - patterns.get(i).getKey();
        for(int j = 0; j < lambda.length; ++j){
            e += lambda[j] * patterns.get(j).getKey() *
                    kernel(patterns.get(j).getValue(), patterns.get(i).getValue());
        }
        eCache[i] = e;
        return e;
    }

    /** 実際の計算処理 */
    private boolean step(int i, int j) {
        if(i == j) return false;
        double fx2 = calcE(j);

        int t1 = patterns.get(i).getKey();
        int t2 = patterns.get(j).getKey();

        double fx1 = calcE(i);

        //基準点2を計算
        double k11 = kernel(patterns.get(i).getValue(), patterns.get(i).getValue());
        double k22 = kernel(patterns.get(j).getValue(), patterns.get(j).getValue());
        double k12 = kernel(patterns.get(i).getValue(), patterns.get(j).getValue());
        double eta = k11 + k22 - 2 * k12;
        if(eta <= 0) return false;
        double newwj = lambda[j] + t2 * (fx1 - fx2) / eta;
        //クリッピング
        double u;
        double v;
        if(t1 == t2){
            u = Math.max(0, lambda[j] + lambda[i] - c);
            v = Math.min(c, lambda[j] + lambda[i]);
        }else{
            u = Math.max(0, lambda[j] - lambda[i]);
            v = Math.min(c, c + lambda[j] - lambda[i]);
        }
        if(u == v) return false;
        newwj = Math.max(u, newwj);
        newwj = Math.min(v, newwj);

        //基準点2から基準点1を計算
        z += Math.abs(lambda[j] - newwj);
        lambda[i] += t1 * t2 * (lambda[j] - newwj);
        lambda[j] = newwj;
        return true;
    }

    /**
    * @param args the command line arguments
    */
    public static void main(String args[]) {
        java.awt.EventQueue.invokeLater(new Runnable() {
            public void run() {
                new ColorCoherenceVecotrLaern().setVisible(true);
            }
        });
    }

    // Variables declaration - do not modify
    private javax.swing.JTextPane tpResut;
    private javax.swing.JTextField txtSearch;
    // End of variables declaration

}