通用圖像識別的神經網絡代碼描述

寫人臉檢測程序的時候順帶寫的，網絡格式是靠讀入一個文件定義的，文件的格式如下：

輸入圖像長 輸入圖像寬 隱層神經元個數 輸出神經元個數
不同網絡結構數量
[連接位置不同的隱層神經元的個數 連接的隱層神經元個數]
[隱層神經元連接的輸入神經元的位置表]

下面是一個例子：

24 28 52 1
3
16 32
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
1 1 1 1 1 1 2 2 2 2 2 2 3 3 3 3 3 3 4 4 4 4 4 4
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
5 5 5 5 5 5 6 6 6 6 6 6 7 7 7 7 7 7 8 8 8 8 8 8
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
9 9 9 9 9 9 10 10 10 10 10 10 11 11 11 11 11 11 12 12 12 12 12 12
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
13 13 13 13 13 13 14 14 14 14 14 14 15 15 15 15 15 15 16 16 16 16 16 16
4 8
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
1 1 1 1 1 1 1 1 1 1 1 1 2 2 2 2 2 2 2 2 2 2 2 2
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
3 3 3 3 3 3 3 3 3 3 3 3 4 4 4 4 4 4 4 4 4 4 4 4
6 12
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1 1
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2 2
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3 3
4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4 4
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5 5
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6
6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6 6

下面是程序代碼：

type

  TSingleExtendedArray = array of extended;
  TDoubleExtendedArray = array of array of extended;

  TSamples = packed record
    Ins: TSingleExtendedArray;
    Outs: TSingleExtendedArray;
  end;

type

  TGraphicBpnn = class
  private
    procedure BackPropagate(t: TSingleExtendedArray; n, m: extended);
    function UpDate(inputs: TSingleExtendedArray): extended;
  public
    samplecounts, TestCounts: longint;
    procedure AddToTrain(Ins, Outs: TSingleExtendedArray);
    procedure AddToTest(Ins, Outs: TSingleExtendedArray);
    procedure SaveToFile(FileName: string);
    procedure LoadFromFile(FileName: string);
    procedure Train(n, m: extended);
    function Init(FileName: string): boolean;
    function Predict(Ins: TSingleExtendedArray): extended;
    function Test: extended;
    destructor Destroy; override;
  private
    nI, nH, nO: longint;
    aI, aH, aO, Output_Deltas, Hidden_Deltas: TSingleExtendedArray;
    wI, wO, cI, cO: TDoubleExtendedArray;
    Connections: array of array of boolean;
    Samples: array of TSamples;
    TestSet: array of TSamples;
  end;

implementation

function TGraphicBpnn.Init(FileName: string): boolean;
var
  i, j, k, fi, fj: longint;
  nIw, nIh, RopMax, RopNum, RopTypes: longint;
  RopMap: array of longint;
begin
  AssignFile(Input, FileName);
  ReSet(Input);
  Readln(Input, nIw, nIh, nH, nO);
  nI := nIw * nIh;
  setlength(aI, nI);
  setlength(aH, nH);
  setlength(aO, nO);
  for i := 0 to nI - 1 do aI[i] := 1;
  for i := 0 to nH - 1 do aH[i] := 1;
  for i := 0 to nO - 1 do aO[i] := 1;

  setlength(wI, nI, nH);
  setlength(wO, nH, nO);
  setlength(cI, nI, nH);
  setlength(cO, nH, nO);
  setlength(Connections, nI, nH);

  for i := 0 to nI - 1 do
    for j := 0 to nH - 1 do
      Connections[i, j] := False;

  Readln(RopTypes); fj := 0;
  for k := 1 to RopTypes do begin
    Readln(RopMax, RopNum);
    setlength(RopMap, nI);
    fi := 0;
    for i := 1 to nIh do begin
      for j := 1 to nIw do begin
        Read(RopMap[fi]);
        Inc(fi);
      end;
      Readln;
    end;
    fi := 0;
    for i := 1 to RopNum do begin
      Inc(fi);
      if fi > RopMax then fi := 1;
      for j := 0 to nI - 1 do
        if RopMap[j] = fi then Connections[j, fj] := true;
      Inc(fj);
    end;
  end;

  setlength(Output_Deltas, nO);
  setlength(Hidden_Deltas, nH);

  randomize;
  for i := 0 to nI - 1 do
    for j := 0 to nH - 1 do begin
      cI[i, j] := 0;
      wI[i, j] := random(40000) / 10000 - 2;
    end;

  for i := 0 to nH - 1 do
    for j := 0 to nO - 1 do begin
      cO[i, j] := 0;
      wO[i, j] := random(40000) / 10000 - 2;
    end;

  setlength(Samples, $100); setlength(TestSet, $100);
  samplecounts := 0; TestCounts := 0;
  CloseFile(Input);
end;

procedure TGraphicBpnn.BackPropagate(t: TSingleExtendedArray; n, m: extended);
var
  i, j, k: Longint;
  Sum, Change: extended;
begin
  for i := 0 to nO - 1 do
    Output_Deltas[i] := aO[i] * (1 - aO[i]) * (t[i] - aO[i]);

  for j := 0 to nH - 1 do begin
    Sum := 0;
    for k := 0 to nO - 1 do
      Sum := Sum + Output_Deltas[k] * wO[j, k];
    Hidden_Deltas[j] := aH[j] * (1 - aH[j]) * Sum;
  end;

  for j := 0 to nH - 1 do
    for k := 0 to nO - 1 do begin
      Change := Output_Deltas[k] * aH[j];
      wO[j, k] := wO[j, k] + n * Change + m * cO[j, k];
      cO[j, k] := Change;
    end;

  for i := 0 to nI - 1 do
    for j := 0 to nH - 1 do
      if Connections[i, j] then begin
        Change := Hidden_Deltas[j] * aI[i];
        wI[i, j] := wI[i, j] + n * Change + m * cI[i, j];
        cI[i, j] := Change;
      end;

end;

function TGraphicBpnn.UpDate(inputs: TSingleExtendedArray): extended;
var
  i, j, k: Longint;
  Sum: extended;
begin
  for i := 0 to nI - 1 do
    aI[i] := Inputs[i];
  for j := 0 to nH - 1 do begin
    Sum := 0;
    for i := 0 to nI - 1 do
      if Connections[i, j] then
        Sum := Sum + aI[i] * wI[i, j];
    aH[j] := 1 / (1 + Exp(-Sum));
  end;
  for k := 0 to nO - 1 do begin
    Sum := 0;
    for j := 0 to nH - 1 do
      Sum := Sum + aH[j] * wO[j, k];
    aO[k] := 1 / (1 + Exp(-Sum));
  end;
  UpDate := aO[0];
end;

procedure TGraphicBpnn.Train(n, m: extended);
var i: Longint;
begin
  for i := 0 to samplecounts - 1 do begin
    UpDate(Samples[i].Ins);
    BackPropagate(Samples[i].Outs, n, m);
  end;
end;

procedure TGraphicBpnn.AddToTrain(Ins, Outs: TSingleExtendedArray);
var i: longint;
begin
  if samplecounts > High(Samples) then setlength(Samples, samplecounts + $100);
  setlength(Samples[samplecounts].Ins, nI);
  setlength(Samples[samplecounts].Outs, nO);
  for i := 0 to nI - 1 do Samples[samplecounts].Ins[i] := Ins[i];
  for i := 0 to nO - 1 do Samples[samplecounts].Outs[i] := Outs[i];
  Inc(samplecounts);
end;

procedure TGraphicBpnn.AddToTest(Ins, Outs: TSingleExtendedArray);
var i: longint;
begin
  if TestCounts > High(TestSet) then setlength(TestSet, TestCounts + $100);
  setlength(TestSet[TestCounts].Ins, nI);
  setlength(TestSet[TestCounts].Outs, nO);
  for i := 0 to nI - 1 do TestSet[TestCounts].Ins[i] := Ins[i];
  for i := 0 to nO - 1 do TestSet[TestCounts].Outs[i] := Outs[i];
  Inc(TestCounts);
end;

procedure TGraphicBpnn.SaveToFile(FileName: string);
var
  i, j, k: longint;
  SaveStream: TMemoryStream;
begin
  SaveStream := TMemoryStream.Create;
  SaveStream.Seek(0, 0);
  for i := 0 to nI - 1 do
    for j := 0 to nH - 1 do begin
      SaveStream.Write(wI[i, j], sizeof(wI[i, j]));
      SaveStream.Write(cI[i, j], sizeof(cI[i, j]));
    end;
  for j := 0 to nH - 1 do
    for k := 0 to nO - 1 do begin
      SaveStream.Write(wO[j, k], sizeof(wO[j, k]));
      SaveStream.Write(cO[j, k], sizeof(cO[j, k]));
    end;
  SaveStream.SaveToFile(FileName);
  SaveStream.Free;
end;

procedure TGraphicBpnn.LoadFromFile(FileName: string);
var
  i, j, k: longint;
  ReadStream: TMemoryStream;
begin
  ReadStream := TMemoryStream.Create;
  ReadStream.LoadFromFile(FileName);
  ReadStream.Seek(0, 0);
  for i := 0 to nI - 1 do
    for j := 0 to nH - 1 do begin
      ReadStream.Read(wI[i, j], sizeof(wI[i, j]));
      ReadStream.Read(cI[i, j], sizeof(cI[i, j]));
    end;
  for j := 0 to nH - 1 do
    for k := 0 to nO - 1 do begin
      ReadStream.Read(wO[j, k], sizeof(wO[j, k]));
      ReadStream.Read(cO[j, k], sizeof(cO[j, k]));
    end;
  ReadStream.Free;
end;

function TGraphicBpnn.Predict(Ins: TSingleExtendedArray): extended;
begin
  try
    Predict := Update(Ins);
  except
    Predict := 0;
  end;
end;

function TGraphicBpnn.Test: extended;
var
  PreRet: extended;
  i, Counts, Ret: longint;
begin
  Counts := 0;
  for i := 0 to TestCounts - 1 do begin
    PreRet := Predict(TestSet[i].Ins);
    if PreRet > 0.5 then Ret := 1 else Ret := 0;
    if Ret = TestSet[i].Outs[0] then Inc(Counts);
  end;
  Result := Counts / TestCounts;
end;

destructor TGraphicBpnn.Destroy;
begin
  setlength(aI, 0);
  setlength(aH, 0);
  setlength(aO, 0);
  setlength(Output_Deltas, 0);
  setlength(Hidden_Deltas, 0);
  setlength(wI, 0, 0);
  setlength(wO, 0, 0);
  setlength(cI, 0, 0);
  setlength(cO, 0, 0);
  setlength(Connections, 0, 0);
  setlength(Samples, 0);
  inherited;
end;