本文主要做了什么
从摄像头读取每一帧的图片,用一些简单的方法将多张图片信息压缩到一份文件中(自定义的视频文件),自定义解码器读取视频文件,并将每帧图片展示成视频
第一步:按照某些算法帧内压缩
常见的视频压缩算法(H264,H265,MP4)过程很复杂,实现的压缩比率也很恐怖(H265可以做到0.5%的压缩率,也就是就算每帧图片加起来有2个GB,合并起来的视频也就10MB),其中压缩算法流程大致如下,我的程序没有细究算法,简单实现了25%的压缩率。
帧内压缩:
- 帧分割:
将原本RGB格式的图像用YUV表示,用YUV是将原本的像素信息转化成亮度和色度信息,由于人眼对色度的变化并不敏感,所以YUV可以在多个像素点之上采用同一数据以实现数据压缩。具体的做法是:将原本图片分成22 / 44 / 88 / 1616的宏块,每个宏块(4*4为例)内按照YUV格式数据采集——记录每个像素格的亮度Y,记录每横向两个像素格的色度U,记录每个宏块左上角像素各的色度V。算法将Y,U,V分别存储,再在接收端分别取出某个宏块对应的数据,恢复成YUV,再恢复成RGB。 - 帧内预测:
邻近的宏块之间可以进行预测,算法思想是由一个宏块,通过某种预测模式,得到一个预测的模块,将实际值和预测值之间的残差进行保存。 - 离散余弦变换(DCT)
对每个块的残差执行DCT变换,算法思想是:图像数据分为细节、纹理和快速变化这类的高频信息,和像整体趋势、平均值和慢速变化这类低频信息;DCT主要保留包含了数据整体特征的低频信息。 - 量化:
由于DCT的结果中浮点数较多,量化将其截断为整数以减少数据量 - 熵编码:
熵编码用于编码多种类型的信息,像文本、图像、音频等信息根据数据的概率分布(如字符、像素、采样值)映射为可变长度的编码。经典哈夫曼树就是一种实现。在此就是将像素值/YUV值根据其概率分布设置不同编码。
帧间压缩:
- 帧间预测:
由于很多帧之间存在冗余,算法首先选择一个参考帧,然后计算参考帧和当前帧之间的运动矢量,由此去除冗余信息 - 运动补偿…
- 残差计算…
- …
我的代码:
- 主要Controller:
@GetMapping("/compressedVideos")
public void getCompressedBytes() throws IOException {
//录制5秒的视频,存在List中
webcam.open();
long startTime = System.currentTimeMillis();
List<BufferedImage> bufferedImages = new ArrayList<>();
while (System.currentTimeMillis() - startTime < 5000) {
BufferedImage image = webcam.getImage();
bufferedImages.add(image);
}
System.out.println("录制结束");
webcam.close();
//调用压缩方法,将结果写入文件中
byte[] bytes = outerCompressionUtils.photosToCompressedBytes(bufferedImages);
File file = new File("压缩中的压缩.dat");
FileOutputStream fos = new FileOutputStream(file);
fos.write(bytes);
fos.close();
System.out.println("持久化结束");
}
- 压缩:
- 工具方法:将rgb转化成YUV
public static int[] rgb2YUV(int rgb) {
int[] rgb1 = photoOps.RGBToInts(rgb);
int red = rgb1[0];
int green = rgb1[1];
int blue = rgb1[2];
int Y = (int) (0.299 * red + 0.587 * green + 0.114 * blue -128); //-128 到 127
int U = (int) (-0.1684 * red - 0.3316 * green + 0.5 * blue);//-128 到 127
int V = (int) (0.5 * red - 0.4187 * green - 0.083 * blue); //-128 到 127
return new int[]{Y, U, V};
}
- 工具方法:一张图片化成YUV
public static byte[] compressToOneChannel(BufferedImage bufferedImage) {
byte[] Ys = new byte[bufferedImage.getWidth() * bufferedImage.getHeight()];
byte[] Us = new byte[bufferedImage.getHeight() * (bufferedImage.getWidth() / 2)];
byte[] Vs = new byte[(bufferedImage.getWidth() / 2) * (bufferedImage.getHeight() / 2)];
int targetYs = 0;
int targetUs = 0;
int targetVs = 0;
/*
这里就是遍历2*2的宏块,将其中对应YUV分别写到YUV的数组中
需要注意的是我犯的一个错误:没有注意到Y和U的遍历过程,导致在解码的时候图片异常
*/
for (int i = 0; i < bufferedImage.getHeight(); i += 2) {
for (int j = 0; j < bufferedImage.getWidth(); j += 2) {
for (int k = 0; k < 2; k++) {
for (int l = 0; l < 2; l++) {
int[] ints = rgb2YUV(bufferedImage.getRGB(j + l, i + k));
int Y = ints[0];
Ys[targetYs] = (byte) (Y);
targetYs++;
}
int[] ints = rgb2YUV(bufferedImage.getRGB(j, i + k));
int U = ints[1];
Us[targetUs] = (byte) (U);
targetUs++;
}
int[] ints = rgb2YUV(bufferedImage.getRGB(j, i));
int V = ints[2];
Vs[targetVs] = (byte) (V);
targetVs++;
}
}
int length1 = Ys.length; //大小估计 : 图片3000*2000 = 6000000 不会超int范围
int length2 = Us.length;
int length3 = Vs.length;
byte[] targetBytes = new byte[4 * 5 + length1 + length2 + length3];
int targetIndex = 0;
//这里是将byte[]开头填充一些用于解码的信息,因为Ys,Us,Vs都是一起传的,需要在包开头标明每个数组长度
//Y区的长度
byte[] bytes1 = intToByte(length1);
for (byte b : bytes1) {
targetBytes[targetIndex] = b;
targetIndex++;
}
//U区长度
byte[] bytes2 = intToByte(length2);
for (byte b : bytes2) {
targetBytes[targetIndex] = b;
targetIndex++;
}
//V区长度
byte[] bytes3 = intToByte(length3);
for (byte b : bytes3) {
targetBytes[targetIndex] = b;
targetIndex++;
}
//图片的高
byte[] bytes4 = intToByte(bufferedImage.getHeight());
for (byte b : bytes4) {
targetBytes[targetIndex] = b;
targetIndex++;
}
//图片的宽
byte[] bytes5 = intToByte(bufferedImage.getWidth());
for (byte b : bytes5) {
targetBytes[targetIndex] = b;
targetIndex++;
}
//传递真实数据
for (byte y : Ys) {
targetBytes[targetIndex] = y;
targetIndex++;
}
for (byte u : Us) {
targetBytes[targetIndex] = u;
targetIndex++;
}
for (byte v : Vs) {
targetBytes[targetIndex] = v;
targetIndex++;
}
return targetBytes;
}
- 工具方法:多张图片化成YUV并压缩
public static byte[] photosToCompressedBytes(List<BufferedImage> bufferedImages) throws IOException {
//数据流中未必要有各种辅助信息,比如各类字段长度,在外规定好算了
//这里每一帧的长度就是:20 + 640 * 480 * 1.75
ByteArrayOutputStream baos = new ByteArrayOutputStream();
//java提供的压缩工具,此输出流将输出的东西压缩输出
//传入的Deflater对象用于控制压缩算法
DeflaterOutputStream dos = new DeflaterOutputStream(baos,new Deflater());
//帧信息添加到压缩流
for (BufferedImage bufferedImage: bufferedImages
) {
byte[] bytes = innerCompressionUtils.compressToOneChannel(bufferedImage);
System.out.println("一帧的长度为:"+bytes.length);
dos.write(bytes);
}
byte[] compressedData = baos.toByteArray();
return compressedData;
}
- 尝试用哈夫曼编码优化
class HuffmanNode implements Comparable<HuffmanNode>{
byte value;
int frequency;
HuffmanNode left;
HuffmanNode right;
public HuffmanNode(byte value,int frequency){
this.value = value;
this.frequency = frequency;
}
@Override
public int compareTo(@NotNull HuffmanNode o) {
return this.frequency - o.frequency;
}
}
public class Huffman {
public static Map<Byte,String> encodingTable;
public static String huffmanEncoding(byte[] originalBytes){
Map<Byte,Integer> frequencyMap = new HashMap<>();
for (byte b: originalBytes
) {
frequencyMap.put(b, frequencyMap.getOrDefault(b,0)+1);
}
PriorityQueue<HuffmanNode> minHeap = new PriorityQueue<>();
for (Map.Entry<Byte, Integer> entry : frequencyMap.entrySet()
) {
minHeap.add(new HuffmanNode(entry.getKey(),entry.getValue()));
}
while (minHeap.size()>1){
HuffmanNode left = minHeap.poll();
HuffmanNode right = minHeap.poll();
HuffmanNode mergeNode = new HuffmanNode((byte)0, left.frequency + right.frequency);
mergeNode.left = left;
mergeNode.right = right;
minHeap.add(mergeNode);
}
encodingTable = new HashMap<>();
HuffmanNode root = minHeap.poll();
buildEncodingTable(root,"",encodingTable);
StringBuilder encodingData = new StringBuilder();
for (Byte b: originalBytes
) {
encodingData.append(encodingTable.get(b));
}
System.out.println("原始数组长度"+originalBytes.length);
System.out.println("哈夫曼后数组长度"+encodingData.length());
return encodingData.toString();
}
public static void buildEncodingTable(HuffmanNode node,String currentCode,Map<Byte,String> encodingMap) {
if (node == null) {
return;
}
if (node.left == null && node.right == null) {
encodingMap.put(node.value, currentCode);
} else {
buildEncodingTable(node.left, currentCode + "0", encodingMap);
buildEncodingTable(node.right, currentCode + "1", encodingMap);
}
}
但其实这里用哈夫曼并不会优化数据量,原因如下:
我传输的数据是-128到127的byte类型,这些byte来自图片的亮度和色度,调试中发现这255个数字出现的频率差不多,全部都在14万到20万之间,两个最小值加起来任然比最大值大,这就意味着这颗哈夫曼树会比较满,类似完全二叉树,于是就无法区分出现频率最高的某个字符。
另外,原本255个数将8位byte全都占满,假如有一个频率很高的元素,我们把较短的0101赋给它,那势必会导致原本以0101开头的元素用8位以上的长度进行表示,而程序中各元素出现频率相近,这就会导致如果有元素用短于8位的编码,其他长于8位编码的元素会导致数据更加庞大。
我在用huffman编码后,数据量一点都没有变,只是由长度为40647865的byte数组变成长度为325182920的字符串,其实就是×8 。怀疑是代码哪里错了…
常见的压缩算法是将DCT变换后的结果进行哈夫曼编码,DCT变换后低频信息和高频信息自然区分开,确实更适合这个熵编码方法
3. 解压:
1. 先将java zip包的压缩过程解压
public static InflaterInputStream inflaterCompressedBytes(byte[] bytes) throws IOException {
//解压数据
ByteArrayInputStream bais = new ByteArrayInputStream(bytes);
InflaterInputStream lis = new InflaterInputStream(bais, new Inflater());
return lis;
}
- 依据压缩时自定义的格式进行对byte数组解析
public static BufferedImage getBfi(byte[] originalBytes) {
//分别先把开头表示各个区长度以及图片宽高的参数取出来
byte one = originalBytes[0];
byte two = originalBytes[1];
byte three = originalBytes[2];
byte four = originalBytes[3];
int Y = ((one & 0xff) << 24) | ((two & 0xff) << 16) | ((three & 0xff) << 8) | (four & 0xff);
byte one2 = originalBytes[4];
byte two2 = originalBytes[5];
byte three2 = originalBytes[6];
byte four2 = originalBytes[7];
int U = ((one2 & 0xff) << 24) | ((two2 & 0xff) << 16) | ((three2 & 0xff) << 8) | (four2 & 0xff);
byte one3 = originalBytes[8];
byte two3 = originalBytes[9];
byte three3 = originalBytes[10];
byte four3 = originalBytes[11];
int V = ((one3 & 0xff) << 24) | ((two3 & 0xff) << 16) | ((three3 & 0xff) << 8) | (four3 & 0xff);
byte one4 = originalBytes[12];
byte two4 = originalBytes[13];
byte three4 = originalBytes[14];
byte four4 = originalBytes[15];
int height = ((one4 & 0xff) << 24) | ((two4 & 0xff) << 16) | ((three4 & 0xff) << 8) | (four4 & 0xff);
byte one5 = originalBytes[16];
byte two5 = originalBytes[17];
byte three5 = originalBytes[18];
byte four5 = originalBytes[19];
int width = ((one5 & 0xff) << 24) | ((two5 & 0xff) << 16) | ((three5 & 0xff) << 8) | (four5 & 0xff);
System.out.println("Y: " + Y);
//将数据读取出来
byte[] Ys = Arrays.copyOfRange(originalBytes, 20, Y + 20);
byte[] Us = Arrays.copyOfRange(originalBytes, Y + 20, Y + U + 20);
byte[] Vs = Arrays.copyOfRange(originalBytes, Y + U + 20, Y + U + V + 20);
BufferedImage bfi = new BufferedImage(width, height, BufferedImage.TYPE_INT_RGB);
int hongW = width / 2;
int hongH = height / 2;
//用YUV数据恢复成RGB,填充到图片的每一个像素
for (int i = 0; i < height - 1; i++) {
for (int j = 0; j < width - 1; j++) {
int H = i / 2;
int W = j / 2;
byte y = Ys[(i / 2 * 2) * width + j / 2 * 4 + (i % 2) * 2 + j % 2];
byte u = Us[H * hongW * 2 + j / 2 * 2 + i % 2];
byte v = Vs[H * hongW + W];
int r = (int) (y + 128 + 1.14075 * (v));
int g = (int) (y + 128 - 0.3455 * (u) - 0.7169 * (v));
int b = (int) (y + 128 + 1.779 * (u));
r = Math.min(255, Math.max(0, r));
g = Math.min(255, Math.max(0, g));
b = Math.min(255, Math.max(0, b));
int color = (r) << 16 | (g) << 8 | b;
if (i < 1 && j < 20) {
bfi.setRGB(j, i, color);
}
}
return bfi;
}
- 简单的播放器(基于Swing)
FileInputStream fileInputStream = new FileInputStream("C:\\Users\\吴松林\\IdeaProjects\\meitu2\\压缩中的压缩.dat");
ByteArrayOutputStream outputStream = new ByteArrayOutputStream(); //此输出流中写入所有信息,最后转出为byte[],类似桶子
byte[] buffer = new byte[1024];
int bytesRead;
while ((bytesRead = fileInputStream.read(buffer))!=-1){
outputStream.write(buffer,0,bytesRead);
}
byte[] data = outputStream.toByteArray();
InflaterInputStream iutputStream1 = utils.inflaterCompressedBytes(data); //解压
BufferedInputStream bis = new BufferedInputStream(iutputStream1);
List<BufferedImage> bufferedImages = new ArrayList<>();
byte[] eachImage = new byte[(int) (20+640*480*1.75)];
int testIndex = 0;
int index;
System.out.println("length: "+eachImage.length);
try {
while ((index = bis.read(eachImage)) != -1) {
System.out.println("本次读取长度:" + index);
testIndex++;
System.out.println("test: " + testIndex);
BufferedImage bfi = utils.getBfi(eachImage);
bufferedImages.add(bfi);
}
}catch (Exception e){
System.out.println("跳过异常,省略最后一张图片");
e.printStackTrace();
}
bis.close();
iutputStream1.close();
outputStream.close();
fileInputStream.close();
JFrame jFrame = new JFrame();
myPanel panel = new myPanel();
jFrame.add(panel);
jFrame.setSize(new Dimension(640,480));
jFrame.setVisible(true);
panel.list = bufferedImages;
while (true){
panel.repaint();
}
}
}
class myPanel extends JPanel{
int index = 0;
List<BufferedImage> list;
@Override
public void paint(Graphics g) {
g.drawImage(list.get(index), 0, 0, null);
if (index < list.size() - 2) {
index++;
}
try {
Thread.sleep(34);
} catch (InterruptedException e) {
throw new RuntimeException(e);
}
}
}
注意:
- zip包在使用时我遇到报:Unexpected end of ZLIB input stream,没找到很合适的解决办法,但发现这个异常是在读取到最后一张图片时才触发,于是我选择舍弃最后一张图
- 这个播放器只用Swing简单写了一个用于测试能否读取文件,很明显我的播放器只能播放我的视频,因为其解码方式和编码方式息息相关,而各种常见的编码方式里的算法又太过复杂。所以这个程序就相当于写着玩而已,和其他视频/播放器难有半点干系。