I have read a lot of articles on analyzing JPEG images on the Internet, and there are more or less problems. The following are the links to the articles I have referred to:
First, the steps of parsing
1. Read the information of the file
Information in the JPEG format is stored in segments (data structures).
The format of the segment is as follows
| name | Bytes | data | illustrate |
|---|---|---|---|
| segment identifier | 1 | FF | start identifier for each new segment |
| segment type | 1 | type encoding (called "marker code") | |
| segment length | 2 | Includes segment content and segment length itself, excluding segment ID and segment type | |
| Column content | ≤65533 bytes |
For the rest of the specific information, please refer to the link below, I will not be a repeater.
Introduction to JPEG tag description
format
It is worth noting that the high bit of a byte is on the left, and there are 8 DC component reset tags in total. Other format descriptions are detailed enough in the second link
Segments that must be read in these segments: SOS, DHT, DQT, SOF, DRI, and others are just icing on the cake. There may
be multiple SOF segments. We need to get the maximum height and width of the image in these segments. , and the horizontal and vertical sampling factors of YCbCr are recorded as Hmax and Vmax respectively, and the start interval in DRI will be used later. It refers to the DC component reset interval, which we record as reset
2. Huffman encoding and decoding
First of all, Huffman coding is divided into DC table (DC) and AC table (AC). They generally have two tables. Which table to use is determined by the corresponding relationship in SOS. Generally, Y corresponds to the first table, and CbCr corresponds to the second table. Two or three tables.
Because it is stipulated that huffman encoding can be up to 16 bits, the code below the maximum value of huffman encoding is 65535,
which is my decoding method, which is the same for DC and AC
int curPos = 16, curCode = 0;
for (int i = 0; i < 16; i++) {
int count = temp[i];//count为二进制位数为i+1的个数
curCode <<= 1; //curCode为当前huffman编码数值
while (count--) {
//一次循环生成一个
uint16_t code=curCode;
uint8_t bit=i+1;//比特位有几位 00为2位
uint8_t weight=temp[curPos];//权重是按照顺序排列的，如比特位为两位的编码有两个，设为00，01，后面权重排列为1，2，则00对应1，01对应2
pair<uint8_t, uint8_t> t1(bit,weight);
//<code,<bit,weight>
pair<uint16_t, pair<uint8_t, uint8_t>> t2(curCode,t1);
table.insert(t2);
curCode++;
curPos++;
}
}
3. Analysis of DC and AC codes
After the SOS segment is the real image compression data, you can choose to read it into the memory at one time, or you can read the data while doing the subsequent analysis steps, I chose the second option. Do a parsing after reading an MCU (I use a cache queue). When encoding pictures, it is necessary to divide the MCU (Minimum Coding Unit). Each MCU is composed of multiple 8×8 matrices. The two-dimensional array is converted into one-dimensional by encoding, so when the read data reaches 64, it is Represents the completion of an 8×8 block parsing until the end of reading 0xFFD9
However, how many 8×8 matrices can be read to resolve an MCU?
The number of 8×8 matrices in the MCU, from the perspective of encoding, the number of 8×8 matrices is Hmax*Vmax, but from the perspective of decoding, because YCbCr has been separated into three tables at this time, Therefore, the number of 8×8 matrices should be the sum of the products of the horizontal and vertical sampling factors of the three components (multiply first, then summation) and record it as SUMmcu, so you need to read SUMmcu 8×8 matrices once (at this time There are three tables of YCbCr, and then convert YCbCr to RGB value by formula)
Well, here we know how many 8×8 matrices to read (in fact, because there is no anti-Zig-Zag encoding, there is still a one-dimensional array of 64 data at this time)
Next, we need to use The Huffman code decoded in the previous step.
The analysis method is as follows:
1. For DC (differential coding) , read the image compression data according to one bit , if the code is found in the Huffman table, and the number of bits is equal , then read the corresponding weight of the code, the weight It represents how many bits to read next as the value of the DC component. Do you think this is over? Also add the differential correction variables (YCbCr has one for each table, so there are 3 in total) .
2. For AC (run-length encoding) , the other parts are the same (there is no differential correction variable). For example, if the DC component read is 0x37, the lower 4 bits (7 here) represent the next 7 The first bit is the value of the AC component, and the upper 4 bits (here 3) represent that there are three 0s before the AC component (there is no need to add the previous one here) . Note that the Huffman table used by DC and AC is different.
3. The next step is to read the AC component in a loop, so when to quit?
There are two conditions, as long as one is met, you can exit
- 63 AC components are read
- The weight of the AC component is 0, and this bit is followed by all 0
For the value read according to the weight (no distinction between DC and AC), if the highest bit (the leftmost) is 0, it is a negative number, otherwise it is a positive number, the judgment code is as follows, curValue is the read value, and curValueLength is the read value How many digits are there in the value
curValue = (curValue >= pow(2, curValueLength - 1) ? curValue : curValue - pow(2, curValueLength) + 1);
There are still two pits in it (if the DC component reset interval reset by DRI is 0, don’t worry about this step)
- Assuming that reset is 332 (this is the interval of my picture), it is separated by 332 MCUs (that is, 332×SUMmcu 8×8 matrix), all differential correction variables need to be set to 0, and when these 332
MCUs After reading, you need to read two bytes (these two bytes are a segment), these two bytes should be exactly 0xFF 0xD0~0xD7, and D0 to D7 appear in order, for example, above If one is 0xFFD0, then the next one must be 0xD1, and after D7, the next one is D0. If it doesn’t match, there is a problem. Also, the two bytes read out are not the compressed data of the picture and do not need to be decoded - If 0xFF00 is read, 00 is ignored
At this point, we have a one-dimensional array with 64 elements
4. Dequantization
We use the quantization table read out before (it is also 64 elements, what a coincidence you say) and multiply it with the corresponding item of the element decoded above, and the dequantization is completed! ! !
5. Anti Zig-Zag change
The encoding method is as follows
The simulation method I use is to convert a one-dimensional array into an 8×8 matrix.
The function is as follows, which is not well written.
double** UnZigZag(int* originArray){
double** table=new double*[ROW];
for(int i=0;i<ROW;i++) table[i]=new double[COL];
int cur=0,x=0,y=0;
bool flag = true;//true是右上 false是左下
while (cur < 64) {
table[y][x] = originArray[cur++];
if (flag) {
x++; y--; }
else {
x--; y++; }
if (x < 0 || y < 0 || x>7 || y>7) flag = !flag;
if (x < 0 && y>7) {
x = 1; y = 7; }
if (x < 0) x = 0;
else if (x > 7) {
x = 7; y += 2; }
if (y < 0) y = 0;
else if (y > 7) {
y = 7; x += 2; }
}
return table;
You can also use another method, manually record an array, write the location, the conversion only needs 4 lines of code
6. Anti-DCT change
That formula is too slow. There is a simplified version of this formula. The formula can be reduced to matrix multiplication, and only one transformation
matrix is needed. The matrix is calculated by the following code
double** JPEGData::createDCTAndIDCTArray(int row){
double** res=new double*[row];
for(int i=0;i<row;i++) res[i]=new double[row];
// cout<<endl;
for(int i=0;i<row;i++){
for(int j=0;j<row;j++){
double t=0;
if(i==0) t=sqrt(1.0/row);
else t=sqrt(2.0/row);
res[i][j]=t*cos(M_PI*(j+0.5)*i/row);
// cout<<res[i][j]<<" ";
}
// cout<<endl;
}
return res;
}
//设返回的矩阵为A
//DCT原理 Y=A*X*A'(X为正变换输入，Y是输出)
//IDCT原理X=A'*Y*A（Y是逆变换输入，X是输出'是转置）
void JPEGData::IDCT(double** originMatrix){
vector<vector<double>> temp(ROW,vector<double>(COL,0));
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=DCTAndIDCTArray[k][i]*originMatrix[k][j];
}
temp[i][j]=sum;
}
}
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=temp[i][k]*DCTAndIDCTArray[k][j];
}
originMatrix[i][j]=sum;
}
}
}
void JPEGData::DCT(double** originMatrix){
vector<vector<double>> temp(ROW,vector<double>(COL,0));
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=DCTAndIDCTArray[i][k]*originMatrix[k][j];
}
temp[i][j]=sum;
}
}
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=temp[i][k]*DCTAndIDCTArray[j][k];
}
originMatrix[i][j]=sum;
}
}
}
8.YCbCr to RGB
The formula is as follows, this is really good so that
R=128+y+1.402 cr
G=128+y-0.71414 cr-0.34414*cb
B=128+y+1.772 *cb
struct RGB{
uint8_t red;
uint8_t green;
uint8_t blue;
};
RGB** JPEGData::YCbCrToRGB(const int* YUV){
RGB **res = new RGB *[ROW * max_v_samp_factor];
int matrixCount = YUV[0] + YUV[1] + YUV[2];
int crCount = 0, cbCount = 0;
//1＝Y, 2＝Cb, 3＝Cr
//式子 scale*x,scale*y
double cb_h_samp_scale=component[1].h_samp_factor*1.0/max_h_samp_factor,
cb_v_samp_scale=component[1].v_samp_factor*1.0/max_v_samp_factor,
cr_h_samp_scale=component[2].h_samp_factor*1.0/max_h_samp_factor,
cr_v_samp_scale=component[2].v_samp_factor*1.0/max_v_samp_factor;
for (int i = 0; i < ROW * max_v_samp_factor; i++)
res[i] = new RGB[COL * max_h_samp_factor];
//此处直接生成rgb值
//注意，此处YCbCr的对应关系与采样因子有关
//这个ycbcr存的是一个MCU，假设YUV为411，那么ycbcr有6个
//这种方式转换不了YUV为420的，因为数组越界了，不过可以加个判断，我懒得改了
// cout<<endl;
for(int j=0;j<ROW * max_v_samp_factor;j++){
for(int k=0;k<COL * max_h_samp_factor;k++){
int yPos = (j / ROW) * component[0].h_samp_factor + (k / COL);
int cbPos = YUV[0] + (int)((k / ROW) * cb_v_samp_scale) + (int)((j / COL) * cb_h_samp_scale);
int crPos = YUV[0] + YUV[1] + (int)((k / ROW) * cr_v_samp_scale) + (int)((j / COL) * cr_h_samp_scale);
double y = ycbcr[yPos][j % ROW][k % COL];
double cb = ycbcr[cbPos][(int)(j * cb_v_samp_scale)][(int)(k * cb_h_samp_scale)];
double cr = ycbcr[crPos][(int)(j * cr_v_samp_scale)][(int)(k * cr_h_samp_scale)];
res[j][k].red =RGBValueLimit(128+y+1.402 *cr);
res[j][k].green =RGBValueLimit(128+y-0.71414*cr-0.34414*cb);
res[j][k].blue =RGBValueLimit(128+y+1.772 *cb);
// 输出当前选择的矩阵
//cout<<dec<<yPos<<" "<<cbPos<<" "<<crPos<<" ";
// cout<<hex<<setw(2)<<setfill('0')<<(int)res[j][k].red
// <<setw(2)<<setfill('0')<<(int)res[j][k].green
// <<setw(2)<<setfill('0')<<(int)res[j][k].blue<<" ";
}
// cout<<endl;
}
// cout<<endl;
return res;
}
renderings
This is a JPEG
and this is a Bitmap
Finally, how to display the picture? , I convert the information to a bitmap and I can see it.
The source code below attaches
Image.h
#pragma once
#define _USE_MATH_DEFINES
#include <cmath>
#include <fstream>
#include <stdint.h>
#include <utility>
#ifndef _IMAGE_
#define _IMAGE_
#include "Util.h"
#include <string>
#include <vector>
#include <iostream>
using namespace std;
NAME_SPACE_START(myUtil)
#define ROW 8
#define COL 8
#define HUFFMAN_DECODE_DEQUE_CACHE 64//单位：位
// #define _DEBUG_
// #define _DEBUGOUT_
#define FREE_VECTOR_LP(vectorName) \
for(auto item : vectorName){
\
for(int i=0;i<ROW;i++)\
delete [] item[i];\
delete [] item; \
}\
vectorName.clear();
//释放二维指针
#define FREE_LP_2(lpName,row) \
for(int i=0;i<row;i++){
\
delete [] lpName[i];\
}\
delete [] lpName;
//段类型
enum JPEGPType{
SOF0 = 0xC0, //帧开始
SOF1 = 0xC1, //帧开始
SOF2 = 0xC2, //帧开始
DHT = 0xC4, //哈夫曼表
SOI = 0xD8, //文件头
EOI = 0xD9, //文件尾
SOS = 0xDA, //扫描行开始
DQT = 0xDB, //定义量化表
DRI = 0xDD, //定义重新开始间隔
APP0 = 0xE0, //定义交换格式和图像识别信息
APP1 = 0xE1, //定义交换格式和图像识别信息
APP2 = 0xE2, //定义交换格式和图像识别信息
COM = 0xFE //注释
};
//将一维数组变为二维数组
double** UnZigZag(int* originArray);
struct RGB{
uint8_t red;
uint8_t green;
uint8_t blue;
};
//SOS
class JPEGScan{
public:
//componentId,<DC,AC>
map<uint8_t,pair<uint8_t,uint8_t>> componentHuffmanMap;
bool Init(fstream& file,uint16_t len);
};
//APP
class JPEGInfo{
public:
uint16_t version;
};
//DHT
class JPEGHuffmanCode{
public:
using iterator = map<uint16_t,pair<uint8_t,uint8_t>>::iterator;
//<code,<bit,weight>
map<uint16_t,pair<uint8_t,uint8_t>> table;
//init huffman table
bool Init(fstream& file,uint16_t len);
//find-true not find-false
bool findKey(const uint16_t& code,const uint8_t& bit,iterator& it);
};
//DQT
//quality table
class JPEGQuality{
public:
uint8_t precision;
uint8_t id;
vector<uint16_t> table;
bool Init(fstream& file,uint16_t len);
};
//SOF segment
class JPEGComponent{
public:
//1＝Y, 2＝Cb, 3＝Cr, 4＝I, 5＝Q
uint8_t colorId;
uint8_t h_samp_factor;
uint8_t v_samp_factor;
uint8_t qualityId;
bool Init(fstream& file,uint16_t len);
};
class JPEGData{
int max_h_samp_factor;//行MCU
int max_v_samp_factor;//列MCU
int width;
int height;
int precision;
bool isYUV411=false;
bool isYUV422=false;
bool isYUV111=false;
uint8_t curDRI=0;//当前重置直流分量标识,这里只取个位方便计算
uint16_t resetInterval=0;//单位是MCU
int preDCValue[3]={
0}; //用于直流差分矫正
//量化表
vector<JPEGQuality> quality;
//huffman码表
vector<JPEGHuffmanCode> dc_huffman;
vector<JPEGHuffmanCode> ac_huffman;
//component每个颜色分量
vector<JPEGComponent> component;
JPEGScan scan;
//vector<int**> deHuffman;
vector<double**> ycbcr;
vector<RGB**> rgb;
double** DCTAndIDCTArray;
streampos pos;
bool EOI{
false};
public:
JPEGData():
max_h_samp_factor(0),
max_v_samp_factor(0),
width(0),
height(0),
precision(0){
DCTAndIDCTArray=createDCTAndIDCTArray(ROW);
}
~JPEGData(){
FREE_LP_2(DCTAndIDCTArray,ROW-1)
// FREE_LP_2(DCTArray,ROW-1)
// FREE_LP_2(IDCTArray,ROW-1)
FREE_VECTOR_LP(rgb)
}
bool readJPEG(const char* filePath);
int getWidth() const {
return width;}
int getHeight() const {
return height;}
vector<RGB**> getRGB() const {
return rgb;}
int getMaxHSampFactor() const {
return max_h_samp_factor;}
int getMaxVSampFactor() const {
return max_v_samp_factor;}
double** createDCTAndIDCTArray(int row);
//double** createIDCTArray(int row);
void DCT(double** originMatrix);
void IDCT(double** originMatrix);
protected:
bool readSOF(fstream& file,uint16_t len);
bool readData(fstream& file);
bool huffmanDecode(fstream& file);
void deQuality(double** originMatrix,int qualityID);
//隔行正负纠正
void PAndNCorrect(double** originMatrix);
RGB** YCbCrToRGB(const int* YUV);
//标记位检查 是否结束,是否重置直流矫正数值，返回要添加的数值
string FlagCkeck(fstream& file,int byteInfo);
uint16_t ReadByte(fstream& file,int len);
uint16_t findHuffmanCodeByBit(fstream& file,int& length,int& pos,string& deque,int curValue,int& curValLen);
};
NAME_SPACE_END()
#endif //!_IMAGE_
Image.cpp
#include "Image.h"
#include "Util.h"
#include <algorithm>
#include <cmath>
#include <exception>
#include <fstream>
#include <stdint.h>
#include <bitset>
#include <stdlib.h>
#include <utility>
#include <cstring>
#include <vector>
#include <iomanip>
NAME_SPACE_START(myUtil)
int RGBValueLimit(double input){
if(input<0) return 0;
else if(input>255) return 255;
// 四舍五入、取整均可
// return (int)(input);
return round(input);
}
void print(double** originMatrix){
cout<<endl;
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
cout<<originMatrix[i][j]<<" ";
}
cout<<endl;
}
cout<<endl;
}
double** UnZigZag(int* originArray){
double** table=new double*[ROW];
for(int i=0;i<ROW;i++) table[i]=new double[COL];
int cur=0,x=0,y=0;
bool flag = true;//true是右上 false是左下
while (cur < 64) {
table[y][x] = originArray[cur++];
if (flag) {
x++; y--; }
else {
x--; y++; }
if (x < 0 || y < 0 || x>7 || y>7) flag = !flag;
if (x < 0 && y>7) {
x = 1; y = 7; }
if (x < 0) x = 0;
else if (x > 7) {
x = 7; y += 2; }
if (y < 0) y = 0;
else if (y > 7) {
y = 7; x += 2; }
}
return table;
}
bool JPEGScan::Init(fstream &file, uint16_t len){
try {
uint8_t count=file.get();
len--;
while(count--){
uint8_t componentId=file.get();
uint8_t table=file.get();
uint8_t dcId=table>>4;
uint8_t acId=table&0x0f;
pair<uint8_t, uint8_t> info1(dcId,acId);
pair<uint8_t, pair<uint8_t, uint8_t>> info2(componentId,info1);
componentHuffmanMap.insert(info2);
}
} catch (...) {
return false;
}
return true;
}
bool JPEGHuffmanCode::Init(fstream &file, uint16_t len){
try{
vector<uint8_t> temp;
while(len--){
int info=file.get();
temp.push_back(info);
}
int curPos = 16, curCode = 0;
for (int i = 0; i < 16; i++) {
int count = temp[i];
curCode <<= 1;
while (count--) {
uint16_t code=curCode;
uint8_t bit=i+1;
uint8_t weight=temp[curPos];
pair<uint8_t, uint8_t> t1(bit,weight);
pair<uint16_t, pair<uint8_t, uint8_t>> t2(curCode,t1);
table.insert(t2);
curCode++;
curPos++;
}
}
}
catch(...){
return false;
}
return true;
}
bool JPEGHuffmanCode::findKey(const uint16_t& code,const uint8_t& bit,iterator& it)
{
it=table.find(code);
if(it==table.end()) return true;
return it->second.first!=bit;
}
bool JPEGQuality::Init(fstream &file, uint16_t len){
try{
int info=file.get();
precision=info>>4;
id=info&0x0f;
len--;
while(len--){
int t=file.get();
table.push_back(t);
}
}
catch(...){
return false;
}
return true;
}
bool JPEGComponent::Init(fstream &file, uint16_t len){
try {
int info1=file.get();
int info2=file.get();
int info3=file.get();
colorId=info1;
h_samp_factor=info2>>4;
v_samp_factor=info2&0x0f;
qualityId=info3;
} catch (...) {
return false;
}
return true;
}
bool JPEGData::readJPEG(const char *filePath){
fstream file(filePath,ios::in|ios::binary);
if(file.fail()) return false;
file.seekg(0,ios::end);
pos = file.tellg();
file.seekg(2,ios::beg);
dc_huffman.resize(2);
ac_huffman.resize(2);
try {
//do read data through using other method
uint16_t pLen=0;
uint16_t pMarker=0xFF;
uint16_t pType=0x00;
while(!file.eof()){
pMarker=file.get();
pType=file.get();
if(pType==EOI) break;
pLen=file.get();
pLen=(pLen<<8)+file.get();
// cout<<hex<<pMarker<<" "<<pType<<" "<<pLen<<endl;
if(pMarker!=0xFF) throw exception();
bool flag=true;
switch (pType) {
case SOF0:
case SOF1:
case SOF2:{
flag=readSOF(file, pLen-2);
break;
}
case DHT:{
JPEGHuffmanCode huf;
int info=file.get();
int tableId=info&0x0f;
// cout<<hex<<info<<" ";
flag=huf.Init(file, pLen-3);
if((info>>4)&1) ac_huffman[tableId]=huf;
else dc_huffman[tableId]=huf;
break;
}
//case SOI:
//case EOI:
case SOS:{
flag=scan.Init(file, pLen-2);
int count=3;
// cout<<endl;
while(count--) file.get();
// cout<<endl;
//正式读取数据
if(!flag) break;
flag=readData(file);
break;
}
case DQT:{
JPEGQuality q;
flag=q.Init(file, pLen-2);
quality.push_back(q);
break;
}
case DRI:{
resetInterval=ReadByte(file, 2);
break;
}
case APP0:
case APP1:
case APP2:
case COM:{
pLen-=2;
while(pLen--){
file.get();
}
break;
}
default:
pLen-=2;
while(pLen--){
file.get();
}
break;
}
if(!flag) throw exception();
// cout<<endl;
}
} catch (...) {
file.close();
return false;
}
file.close();
return true;
}
bool JPEGData::readSOF(fstream& file,uint16_t len){
try {
precision=file.get();
height=max(height,(int)ReadByte(file, 2));
width=max(width,(int)ReadByte(file, 2));
int count=ReadByte(file, 1);
if(count!=3) return false;
len-=6;
component.resize(count);
for(int i=0;i<count;i++){
JPEGComponent com;
com.Init(file, len/3);
max_h_samp_factor=max(max_h_samp_factor,(int)com.h_samp_factor);
max_v_samp_factor=max(max_v_samp_factor,(int)com.v_samp_factor);
component[i]=com;
}
if((component[0].h_samp_factor*component[0].v_samp_factor)
/(component[1].h_samp_factor*component[1].v_samp_factor)==4){
isYUV411=true;
}
else if((component[0].h_samp_factor*component[0].v_samp_factor)
/(component[1].h_samp_factor*component[1].v_samp_factor)==2){
isYUV422=true;
}
else if((component[0].h_samp_factor*component[0].v_samp_factor)
/(component[1].h_samp_factor*component[1].v_samp_factor)==1){
isYUV111=true;
}
} catch (...) {
return false;
}
return true;
}
bool JPEGData::readData(fstream& file){
bool flag=true;
try{
//使用huffman表来解出RLE编码，接着转回长度为64的矩阵
flag=huffmanDecode(file);
if(!flag) return false;
//反量化，即上面的64矩阵×对应位置的量化表
//flag=deQuantity();
//if(!flag) return false;
//反zig-zag排序
//flag=deZSort();
//if(!flag) return false;
//反离散余弦变换
//if(!flag) return false;
//YCbCr转RGB
//if(!flag) return false;
}
catch(...){
return false;
}
return true;
}
bool JPEGData::huffmanDecode(fstream& file){
try {
//原图像一个MCU有多少8*8矩阵（此时是YCbCr还没有分开）
//int MCUBlockCount=max_h_samp_factor*max_v_samp_factor;
//顺序YCbCr
int YUV[]={
component[0].h_samp_factor*component[0].v_samp_factor,
component[1].h_samp_factor*component[1].v_samp_factor,
component[2].h_samp_factor*component[2].v_samp_factor};
int curMCUCount=1; //当前MCU数量
int curValueLength=0; //当前值有多少位
int curValue=0; //当前的值
int curBitDequeLength=8;//当前curBitDeque长度
int curBitPos=0; //当前string读取到第几位
int restart=resetInterval;//直流分量重置
string curBitDeque=""; //用来存储读出来的2进制数
//一次循环解析一个MCU
curBitDeque.append(bitset<8>(file.get()).to_string());
curBitDequeLength=8;
// cout<<curBitDeque;
while(!EOI||(pos-file.tellg())!=2){
// cout<<endl;
int count=1;
for(int i=0;i<3;i++){
for(int j=0;j<YUV[i];j++){
// cout<<count++<<" ";
int matrix[64]={
0};
int valCount=0;
uint8_t dcID = scan.componentHuffmanMap[component[i].colorId].first;
uint8_t acID = scan.componentHuffmanMap[component[i].colorId].second;
int qualityId=component[i].qualityId;
if(qualityId>=quality.size()) qualityId=0;
// cout<<endl;
while(valCount<64){
//用curBitDeque和curBit去找权重，curValue作为当前键值
JPEGHuffmanCode::iterator it;
JPEGHuffmanCode &huffman = valCount==0?dc_huffman[dcID]:ac_huffman[acID];
while(curValueLength<=16&&huffman.findKey(curValue,curValueLength,it)){
curValue=findHuffmanCodeByBit(file,curBitDequeLength,curBitPos,curBitDeque,curValue,curValueLength);
}
if(curValueLength>16)
return true;
#ifdef _DEBUGOUT_
//cout<<dec<<" "<<curBitPos<<" "<<curBitDequeLength<<" ";
cout<<"key="<<hex<<curValue<<" len="<<curValueLength<<endl;
#endif
//已经找到了权重和位宽
uint8_t weight,zeroCount=0;
if(valCount==0) weight = it->second.second;
else {
weight = it->second.second & 0x0f; zeroCount = it->second.second >> 4;}
curValue=0;//这里变为dc或ac值
curValueLength=0;
if(valCount!=0&&weight==0&&zeroCount==0) break;//后面全是0
// 读取真正的值
for(int k=0;k<weight;k++){
curValue=findHuffmanCodeByBit(file,curBitDequeLength,curBitPos,curBitDeque,curValue,curValueLength);
}
curValue = (curValue >= pow(2, curValueLength - 1) ? curValue : curValue - pow(2, curValueLength) + 1);
// cout<<curValue<<endl;
int writeValue=valCount==0?(preDCValue[i]+=curValue):curValue;
valCount+=zeroCount;
writeValue*=quality[qualityId].table[valCount];//反量化
matrix[valCount]=writeValue;
curValue=0;
curValueLength=0;
valCount++;
}
double** tempZ = UnZigZag(matrix);//反zig-zag编码
//反量化，在反zig-zag编码前后差别，前面：RGB数值与编辑器比偏小，反之偏大，这也与最后取整时的方式有关
// deQuality(tempZ,qualityId);
// print(tempZ);
//隔行正负纠正，有的博客说了，但是没感觉有啥帮助
// PAndNCorrect(tempZ);
IDCT(tempZ); //dct逆变换
ycbcr.push_back(tempZ);
#ifdef _DEBUG_
for(int k=0;k<ROW;k++){
for(int l=0;l<COL;l++){
cout.width(3);
cout<<dec<<tempZ[k][j]<<" ";
}
cout<<endl;
}
cout<<endl;
#endif
}
}
// if(count!=6){
// cout<<" ";
// }
RGB** lpRGB = YCbCrToRGB(YUV);
FREE_VECTOR_LP(ycbcr)
rgb.push_back(lpRGB);
// 直流分量重置间隔不为0的
if(restart>0){
resetInterval--;
if(resetInterval==0){
resetInterval=restart;
curDRI+=1;
curDRI&=0x7;
//需要在此处读取两字节信息，看是否为重置标识
file.get();
if(file.get()==0xD9) EOI=true;
curBitPos=curBitDequeLength;
preDCValue[0]=0;
preDCValue[1]=0;
preDCValue[2]=0;
}
}
// cout<<"curMCUCount="<<dec<<curMCUCount++<<" pos="<<pos<<"/"<<file.tellg()<<" "<<file.tellg()*100.0/pos<<"%\n";
if(pos-file.tellg()==2) break;
}
cout<<"\nsuccessfully\n";
} catch (exception ex) {
cout<<ex.what();
return false;
}
return true;
}
RGB** JPEGData::YCbCrToRGB(const int* YUV){
RGB **res = new RGB *[ROW * max_v_samp_factor];
int matrixCount = YUV[0] + YUV[1] + YUV[2];
int crCount = 0, cbCount = 0;
//1＝Y, 2＝Cb, 3＝Cr
//式子 scale*x,scale*y
double cb_h_samp_scale=component[1].h_samp_factor*1.0/max_h_samp_factor,
cb_v_samp_scale=component[1].v_samp_factor*1.0/max_v_samp_factor,
cr_h_samp_scale=component[2].h_samp_factor*1.0/max_h_samp_factor,
cr_v_samp_scale=component[2].v_samp_factor*1.0/max_v_samp_factor;
for (int i = 0; i < ROW * max_v_samp_factor; i++)
res[i] = new RGB[COL * max_h_samp_factor];
//此处直接生成rgb值
//注意，此处YCbCr的对应关系与采样因子有关
// cout<<endl;
for(int j=0;j<ROW * max_v_samp_factor;j++){
for(int k=0;k<COL * max_h_samp_factor;k++){
int yPos = (j / ROW) * component[0].h_samp_factor + (k / COL);
int cbPos = YUV[0] + (int)((k / ROW) * cb_v_samp_scale) + (int)((j / COL) * cb_h_samp_scale);
int crPos = YUV[0] + YUV[1] + (int)((k / ROW) * cr_v_samp_scale) + (int)((j / COL) * cr_h_samp_scale);
double y = ycbcr[yPos][j % ROW][k % COL];
double cb = ycbcr[cbPos][(int)(j * cb_v_samp_scale)][(int)(k * cb_h_samp_scale)];
double cr = ycbcr[crPos][(int)(j * cr_v_samp_scale)][(int)(k * cr_h_samp_scale)];
res[j][k].red =RGBValueLimit(128+y+1.402 *cr);
res[j][k].green =RGBValueLimit(128+y-0.71414*cr-0.34414*cb);
res[j][k].blue =RGBValueLimit(128+y+1.772 *cb);
// 输出当前选择的矩阵
//cout<<dec<<yPos<<" "<<cbPos<<" "<<crPos<<" ";
// cout<<hex<<setw(2)<<setfill('0')<<(int)res[j][k].red
// <<setw(2)<<setfill('0')<<(int)res[j][k].green
// <<setw(2)<<setfill('0')<<(int)res[j][k].blue<<" ";
}
// cout<<endl;
}
// cout<<endl;
return res;
}
double** JPEGData::createDCTAndIDCTArray(int row){
double** res=new double*[row];
for(int i=0;i<row;i++) res[i]=new double[row];
// cout<<endl;
for(int i=0;i<row;i++){
for(int j=0;j<row;j++){
double t=0;
if(i==0) t=sqrt(1.0/row);
else t=sqrt(2.0/row);
res[i][j]=t*cos(M_PI*(j+0.5)*i/row);
// cout<<res[i][j]<<" ";
}
// cout<<endl;
}
return res;
}
void JPEGData::DCT(double** originMatrix){
//原理 Y=A*X*A'
vector<vector<double>> temp(ROW,vector<double>(COL,0));
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=DCTAndIDCTArray[i][k]*originMatrix[k][j];
}
temp[i][j]=sum;
}
}
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=temp[i][k]*DCTAndIDCTArray[j][k];
}
originMatrix[i][j]=sum;
}
}
}
void JPEGData::IDCT(double** originMatrix){
//原理X=A'*Y*A
vector<vector<double>> temp(ROW,vector<double>(COL,0));
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=DCTAndIDCTArray[k][i]*originMatrix[k][j];
}
temp[i][j]=sum;
}
}
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
double sum=0;
for(int k=0;k<COL;k++){
sum+=temp[i][k]*DCTAndIDCTArray[k][j];
}
originMatrix[i][j]=sum;
}
}
}
void JPEGData::deQuality(double** originMatrix,int qualityID){
for(int i=0;i<ROW;i++){
for(int j=0;j<COL;j++){
originMatrix[i][j]*=quality[qualityID].table[i*ROW+j];
}
}
}
void JPEGData::PAndNCorrect(double** originMatrix){
for(int i=0;i<ROW;i++)
if(i%2==1)
for(int j=0;j<COL;j++)
originMatrix[i][j]=-originMatrix[i][j];
}
string JPEGData::FlagCkeck(fstream& file,int byteInfo){
if(byteInfo==0xff){
uint8_t info=file.get();
string res=bitset<8>(0xFF).to_string();
if(info==0xD9) {
EOI=true;return "false";}
else if(info==0x00) return res;
return res + bitset<8>(info).to_string();
}
return bitset<8>(byteInfo).to_string();
}
uint16_t JPEGData::ReadByte(fstream& file,int len){
uint16_t res=file.get();
if(len!=1){
res=(res<<8)+(uint8_t)file.get();
}
return res;
}
uint16_t JPEGData::findHuffmanCodeByBit(fstream& file,int& length,int& pos,string& deque,int curValue,int& curValLen){
if(pos==length&&length>=HUFFMAN_DECODE_DEQUE_CACHE){
//达到最大缓存
deque = deque.substr(pos);
int info=file.get();
string res=FlagCkeck(file,info);
string str=bitset<8>(info).to_string();
if(res=="false") res=bitset<8>(file.get()).to_string();
deque.append(res);
length = deque.length();
pos = 0;
}
else if(length==0 || pos>=length){
if(length==0){
deque="";
pos=0;
}
int info=file.get();
string res=FlagCkeck(file,info);
string str=bitset<8>(info).to_string();
if(res=="false") res=bitset<8>(file.get()).to_string();
deque.append(res);
length+=8;
}
curValue = (curValue << 1) + (uint8_t)(deque.at(pos++) - '0');
curValLen++;
return curValue;
}
NAME_SPACE_END()
BmpEncoder.h
#pragma once
#include <stdio.h>
#include <iostream>
#include "Image.h"
using namespace myUtil;
/* Bitmap Header, 54 Bytes */
static
unsigned char BmpHeader[54] =
{
0x42, 0x4D, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x36, 0x00, 0x00, 0x00, 0x28, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00, 0x1B, 0x01, 0x00, 0x00, 0x01, 0x00, 0x18, 0x00, 0x00, 0x00,
0x00, 0x00, 0x60, 0xCD, 0x04, 0x00, 0x23, 0x2E, 0x00, 0x00, 0x23, 0x2E, 0x00, 0x00, 0x00, 0x00,
0x00, 0x00, 0x00, 0x00, 0x00, 0x00
};
void SetBitmapInfo(unsigned int size, int height, int width)
{
for (int i = 0; i < 4; i++)
{
// size of image ( header + data )
BmpHeader[2 + i] = size & 0xff;
size >>= 8;
// width of image
BmpHeader[18 + i] = width & 0xff;
width >>= 8;
// height of image
BmpHeader[22 + i] = height & 0xff;
height >>= 8;
}
}
/* BGR format 这是我粘来的改了映射部分代码 */
unsigned char *Encoder(const vector<RGB**>& buf, int height, int width, int mcu_height, int mcu_width, int &size)
{
uint8_t *bitmap = nullptr;
int rowSize = (24 * width + 31) / 32 * 4;
// compute the size of total bytes of image
size = rowSize * height + 54; // data size + header size
bitmap = new uint8_t [ size ];
// set the header info
SetBitmapInfo(size, height, width);
// fill the header area
for (int i = 0; i < 54; i++)
{
bitmap[i] = BmpHeader[i];
}
// fill the data area
for (int i = 0; i < height; i++)
{
// compute the offset of destination bitmap and source image
int idx = height - 1 - i;
int offsetDst = idx * rowSize + 54; // 54 means the header length
// int offsetSrc = i * width;
int offsetHeight = (int)floor(i*1.0/mcu_height)*(int)ceil(width*1.0/mcu_width);
// fill data
for (int j = 0; j < width * 3; j++)
{
int pos=(j/3)/mcu_width+offsetHeight;
if(pos>=buf.size()) pos=buf.size()-1;
RGB temp=buf[pos][i%mcu_height][(j/3)%mcu_height];
if(j%3==0) bitmap[offsetDst + j] = temp.blue;
else if(j%3==1) bitmap[offsetDst + j] = temp.green;
else if(j%3==2) bitmap[offsetDst + j] = temp.red;
// cout<<dec<<pos<<" ";
}
// fill 0x0, this part can be ignored
for (int j = width * 3; j < rowSize; j++)
{
bitmap[offsetDst +j] = 0x0;
}
}
return bitmap;
}
/* Save to file */
void Write(const char *fileName, uint8_t *buf, int &size)
{
FILE *fp = fopen(fileName, "wb+");
fwrite(buf, 1, size, fp);
fclose(fp);
}
main program
#include <algorithm>
#include <cctype>
#include <fstream>
#include <iostream>
#include <locale>
#include <sstream>
#include <stdlib.h>
#include "Image.h"
#include "BmpEncoder.h"
using namespace std;
using namespace myUtil;
// void print(double** input){
// cout<<endl;
// for(int i=0;i<8;i++){
// for(int j=0;j<8;j++){
// cout<<input[i][j]<<" ";
// }
// cout<<endl;
// }
// cout<<endl;
// }
int main(){
string str="../img/Image/3.jpg";
JPEGData data;
clock_t startTime=clock();
data.readJPEG(str.c_str());
int size;
unsigned char *bitmap = Encoder(data.getRGB(), data.getHeight(), data.getWidth(),
8*data.getMaxHSampFactor(),
8*data.getMaxVSampFactor(), size);
Write("out.bmp", bitmap, size);
cout<<dec<<clock()-startTime<<"ms"<<endl;
// DCT正反变换测试
// JPEGData data;
// double** arr=new double*[8];
// for(int i=0;i<8;i++){
// arr[i]=new double[8];
// for(int j=0;j<8;j++){
// arr[i][j]=(int)(rand()%100);
// }
// }
// print(arr);
// data.DCT(arr);
// print(arr);
// data.IDCT(arr);
// print(arr);
// FREE_LP_2(arr,8)
return 0;
}
Project environment gcc 7.3.0 tool CMake, source link