基于OpenCV.js编写一个文档扫描Web应用

OpenCV是一个开源的计算机视觉和机器学习软件库。它为我们解决不同的任务提供了各种图像处理方法。

在本文中，我们将使用OpenCV的JavaScript版本编写一个文档扫描Web应用。首先编写了一个JavaScript库，然后与Dynamsoft Document Viewer一起使用，实现一个具有实时检测、自动捕获和文档多边形编辑等功能的文档扫描应用。

演示视频

新建项目

使用Vite创建一个新的TypeScript项目：

bash 复制代码

npm create vite@latest OpenCVDocumentScanner -- --template vanilla-ts

包含OpenCV

在用于开发测试的index.html中，包含OpenCV。

html 复制代码

<script type="text/javascript">
  var Module = {
    // https://emscripten.org/docs/api_reference/module.html#Module.onRuntimeInitialized
    onRuntimeInitialized() {
      document.getElementById('status').innerHTML = 'OpenCV.js is ready.';
    }
  };
</script>
<script async src="https://docs.opencv.org/4.8.0/opencv.js" type="text/javascript"></script>

文档扫描类

接下来，创建一个document-scanner.ts文件。

ts 复制代码

export class DocumentScanner {
  private cv:any;
  constructor() {
    if (!("cv" in window)) {
      throw new Error("OpenCV not found");
    }else{
      this.cv = window["cv"];
    }
  }
}

在构造函数中，添加一个名为cv的属性，以使用cv命名空间来执行相关函数。

检测文档边界

接下来，添加detect方法以返回文档的四个角点。

ts 复制代码

export interface Point {
  x: number;
  y: number;
}

export class DocumentScanner {
  detect(source:HTMLImageElement|HTMLCanvasElement):Point[]{}
}

然后，让我们分步骤实现该方法。

从img或canvas读取图像为OpenCV用的mat类型。
js 复制代码
```
let cv = this.cv;
const img = cv.imread(source);
```

将图像转换为灰度。

js 复制代码

const gray = new cv.Mat();
cv.cvtColor(img, gray, cv.COLOR_RGBA2GRAY);

执行高斯模糊以去除噪点。

js 复制代码

const blur = new cv.Mat();
cv.GaussianBlur(gray,blur,new cv.Size(5, 5),0,0,cv.BORDER_DEFAULT);

执行threshold操作以获得二值化的图像。

js 复制代码

const thresh = new cv.Mat();
cv.threshold(blur,thresh,0,255,cv.THRESH_BINARY + cv.THRESH_OTSU);

寻找轮廓并将最大的轮廓作为候选文档。

js 复制代码

let contours = new cv.MatVector();
let hierarchy = new cv.Mat();
cv.findContours(thresh,contours,hierarchy,cv.RETR_CCOMP,
  cv.CHAIN_APPROX_SIMPLE);

let maxArea = 0;
let maxContourIndex = -1;
for (let i = 0; i < contours.size(); ++i) {
  let contourArea = cv.contourArea(contours.get(i));
  if (contourArea > maxArea) {
    maxArea = contourArea;
    maxContourIndex = i;
  }
}

const maxContour = contours.get(maxContourIndex);

通过计算点到中心点的距离来获得角点。将它们按顺时针方向存储在一个数组中。

js 复制代码

const points = this.getCornerPoints(maxContour);

所需的函数：

js 复制代码

distance(p1:Point, p2:Point) {
  return Math.hypot(p1.x - p2.x, p1.y - p2.y);
}

getCornerPoints(contour:any):Point[] {
  let cv = this.cv;
  let points:Point[] = [];
  let rect = cv.minAreaRect(contour);
  const center = rect.center;

  let topLeftPoint;
  let topLeftDistance = 0;

  let topRightPoint;
  let topRightDistance = 0;

  let bottomLeftPoint;
  let bottomLeftDistance = 0;

  let bottomRightPoint;
  let bottomRightDistance = 0;

  for (let i = 0; i < contour.data32S.length; i += 2) {
    const point = { x: contour.data32S[i], y: contour.data32S[i + 1] };
    const distance = this.distance(point, center);
    if (point.x < center.x && point.y < center.y) {
      if (distance > topLeftDistance) {
        topLeftPoint = point;
        topLeftDistance = distance;
      }
    } else if (point.x > center.x && point.y < center.y) {
      if (distance > topRightDistance) {
        topRightPoint = point;
        topRightDistance = distance;
      }
    } else if (point.x < center.x && point.y > center.y) {
      if (distance > bottomLeftDistance) {
        bottomLeftPoint = point;
        bottomLeftDistance = distance;
      }
    } else if (point.x > center.x && point.y > center.y) {
      if (distance > bottomRightDistance) {
        bottomRightPoint = point;
        bottomRightDistance = distance;
      }
    }
  }
  points.push(topLeftPoint as Point);
  points.push(topRightPoint as Point);
  points.push(bottomRightPoint as Point);
  points.push(bottomLeftPoint as Point);
  return points;
}

释放内存并返回角点。

js 复制代码

img.delete();
gray.delete();
blur.delete();
thresh.delete();
contours.delete();
hierarchy.delete();
return points;

更正文档图像

接下来，添加一个crop方法来运行透视变换，以获得一个矫正好的图像。

js 复制代码

crop(source:HTMLImageElement|HTMLCanvasElement,points?:Point[],width?:number,height?:number):HTMLCanvasElement{}

让我们分步骤讨论实现细节。

如果参数中没有预先提供角点，运行detect以获取文档的角点。
js 复制代码
```
if (!points) {
  points = this.detect(source);
}
```

确定输出图像的宽度和高度。

如果用户未指定宽度和高度，则使用最长的垂直边提供高度，使用最长的水平边提供宽度。

js 复制代码

if (!width) {
  width = Math.max(this.distance(points[0],points[1]),this.distance(points[2],points[3]));
}
if (!height) {
  height = Math.max(this.distance(points[0],points[3]),this.distance(points[1],points[2]));
}

创建一个转换矩阵。

js 复制代码

let srcTri = cv.matFromArray(4, 1, cv.CV_32FC2, [
  points[0].x,
  points[0].y,
  points[1].x,
  points[1].y,
  points[3].x,
  points[3].y,
  points[2].x,
  points[2].y,
]);

let dstTri = cv.matFromArray(4, 1, cv.CV_32FC2, [
  0,
  0,
  width,
  0,
  0,
  height,
  width,
  height,
]);

let M = cv.getPerspectiveTransform(srcTri, dstTri);

执行透视转换。

js 复制代码

let warpedDst = new cv.Mat();
let dsize = new cv.Size(width, height);
cv.warpPerspective(img,warpedDst,M,dsize,cv.INTER_LINEAR,
  cv.BORDER_CONSTANT,
  new cv.Scalar()
);

将结果绘制到canvas元素中。

js 复制代码

const canvas = document.createElement("canvas");
cv.imshow(canvas, warpedDst);

释放内存并返回canvas。

js 复制代码

img.delete()
warpedDst.delete()
return canvas;

使用文档扫描类

我们现在可以使用这个类来获取裁剪后的文档图像。

js 复制代码

const img = document.getElementById("photo");
const documentScanner = new DocumentScanner();
const points = documentScanner.detect(img);
const canvas = documentScanner.crop(img,points);
document.getElementById("output").src = canvas.toDataURL();

例子：

基于Dynamsoft Document Viewer的文档扫描Web应用

Dynamsoft Document Viewer为文档扫描过程提供了多个组件。我们可以使用它的Capture Viewer实现带有相机控制的实时文档扫描功能，并使用它的Perspective Viewer编辑检测到的文档多边形。

实时扫描

Dynamsoft Document Viewer允许通过定义handler来集成第三方文档检测库。可以在此处找到使用Dynamsoft Document Normalizer的示例。

我们将编写一个handler，以便使用刚编写的OpenCV的文档检测类和Capture Viewer来扫描文档。Capture Viewer可以打开相机，绘制检测到的文档边界，并在检测结果稳定的时候自动拍照。

定义Handler

使用以下模板创建一个名为dynamsoft-document-viewer-handler.ts的新文件：

ts 复制代码

import { DocumentScanner } from ".";

let DDV;
//allows setting the DDV namespace. It is needed if Dynamsoft Document Viewer (DDV) is installed with NPM.
export function setDDV(DocumentViewer:any) {
  DDV = DocumentViewer;
}
if ((window as any)["Dynamsoft"]) {
  const Dynamsoft = (window as any)["Dynamsoft"];
  DDV = Dynamsoft.DDV;
}
// Inherit DocumentDetect class
export class OpenCVDocumentDetectHandler extends DDV.DocumentDetect {
  private documentScanner:DocumentScanner;
  constructor(documentScanner:DocumentScanner){
    super();
    this.documentScanner = documentScanner;
  }
  // Rewrite the detect method
  async detect(image:any, detectConfig:any) {}
}

如果图像的高度大于720像素，缩小图像以提高性能。

ts 复制代码

async detect(image:any, detectConfig:any) {
  let width = image.width;
  let height = image.height;

  let ratio = 1;
  let data;
  const thresh = 720
  if (height > thresh) {
    ratio = height / thresh;
    height = thresh;
    width = Math.floor(width / ratio);
    data = this.compress(image.data, image.width, image.height, width, height);
  } else {
    data = image.data.slice(0);
  }
}
compress(
  imageData:any,
  imageWidth:number,
  imageHeight:number,
  newWidth:number,
  newHeight:number,
) {
  let source:any = null;
  try {
    source = new Uint8ClampedArray(imageData);
  } catch (error) {
    source = new Uint8Array(imageData);
  }

  const scaleW = newWidth / imageWidth;
  const scaleH = newHeight / imageHeight;
  const targetSize = newWidth * newHeight * 4;
  const targetMemory = new ArrayBuffer(targetSize);
  let distData = null;

  try {
    distData = new Uint8ClampedArray(targetMemory, 0, targetSize);
  } catch (error) {
    distData = new Uint8Array(targetMemory, 0, targetSize);
  }

  const filter = (distCol:number, distRow:number) => {
    const srcCol = Math.min(imageWidth - 1, distCol / scaleW);
    const srcRow = Math.min(imageHeight - 1, distRow / scaleH);
    const intCol = Math.floor(srcCol);
    const intRow = Math.floor(srcRow);

    let distI = (distRow * newWidth) + distCol;
    let srcI = (intRow * imageWidth) + intCol;

    distI *= 4;
    srcI *= 4;

    for (let j = 0; j <= 3; j += 1) {
      distData[distI + j] = source[srcI + j];
    }
  };

  for (let col = 0; col < newWidth; col += 1) {
    for (let row = 0; row < newHeight; row += 1) {
      filter(col, row);
    }
  }

  return distData;
}

将图像数据绘制到canvas上以供OpenCV使用。

ts 复制代码

const canvas = document.createElement('canvas');
export class OpenCVDocumentDetectHandler extends DDV.DocumentDetect {
  private documentScanner:DocumentScanner;
  constructor(documentScanner:DocumentScanner){
    super();
    this.documentScanner = documentScanner;
  }
  // Rewrite the detect method
  async detect(image:any, detectConfig:any) {
    const quad:any[] = [];
    try {
      //...
      canvas.width = width;
      canvas.height = height;
      const context = canvas.getContext('2d');
      if (context) {
        const imageData = context.createImageData(width, height);
        var pixelData = imageData.data;
        var inputArrayBuffer = data;
        var inputUint8Array = new Uint8Array(inputArrayBuffer);
        pixelData.set(inputUint8Array);
        context.putImageData(imageData, 0, 0);
        const points = this.documentScanner.detect(canvas);
        points.forEach(p => {
          quad.push([p.x * ratio, p.y * ratio]);
        });
      }
    } catch (error) {
      console.log(error);
      quad.push([0, 0]);
      quad.push([0, 0]);
      quad.push([0, 0]);
      quad.push([0, 0]);
    }
  }
}

后处理并返回结果。使用processDetectResult处理结果，它能检查检测到的多边形的形状是否接近文档的形状。

ts 复制代码

const result:any = {
  location: quad,
  width: image.width,
  height: image.height,
  config: detectConfig,
};
const ret = this.processDetectResult(result);
return Promise.resolve(ret);

使用Capture Viewer捕获文档图像

在index.html中，包括Dynamsoft Document Viewer的库，并为Capture Viewer添加一个容器。

html 复制代码

<script src="https://cdn.jsdelivr.net/npm/dynamsoft-document-viewer@1.0.0/dist/ddv.js"></script>
<link rel="stylesheet" href="https://cdn.jsdelivr.net/npm/dynamsoft-document-viewer@1.0.0/dist/ddv.css">
<style>
.fullscreen {
  display: none;
  position: fixed;
  top: 0;
  left: 0;
  width: 100%;
  height: 100%;
}
</style>
<div id="captureViewer" class="fullscreen"></div>

使用以下代码进行设置。可以在这里申请自己的证书。

js 复制代码

let docManager;
let doc;
let captureViewer;
async function initDDV(){
  await Dynamsoft.DDV.setConfig({
      license: "DLS2eyJoYW5kc2hha2VDb2RlIjoiMjAwMDAxLTE2NDk4Mjk3OTI2MzUiLCJvcmdhbml6YXRpb25JRCI6IjIwMDAwMSIsInNlc3Npb25QYXNzd29yZCI6IndTcGR6Vm05WDJrcEQ5YUoifQ==", // Public trial license which is valid for 24 hours
      engineResourcePath: "https://cdn.jsdelivr.net/npm/dynamsoft-document-viewer@1.0.0/dist/engine", // Lead to a folder containing the distributed WASM files
  });
  docManager = Dynamsoft.DDV.documentManager;
  doc = docManager.createDocument({
    name: "doc"
  });
  captureViewer = new Dynamsoft.DDV.CaptureViewer({
    container: document.getElementById("captureViewer"),
    viewerConfig: {
      enableAutoCapture: true,
      enableAutoDetect: true
    }
  });
  captureViewer.openDocument(doc.uid);
}

这里还创建了一个document对象，Capture Viewer等组件会显示和操作这个对象里的图片。

将文档检测的Handler设置为我们刚刚创建的版本。

js 复制代码

let documentScanner = new DocumentScanner();
const detectHandler = new OpenCVDocumentDetectHandler(documentScanner);
Dynamsoft.DDV.setProcessingHandler("documentBoundariesDetect", detectHandler);

创建一个函数来获取裁剪后的文档图像。

js 复制代码

function getCroppedImage(points){
  let canvas = documentScanner.crop(document.getElementById("photoRaw"),points);
  document.getElementById("output").src = canvas.toDataURL();
}

为Capture Viewer的captured事件添加侦听器。拍得照片时，将其加载到img元素中，并获得裁剪后的图像。

js 复制代码

captureViewer.on("captured", async (e) => {
  // Stop video stream and hide capture viewer's container
  captureViewer.stop();
  document.getElementById("captureViewer").style.display = "none";
  const pageData =  await captureViewer.currentDocument.getPageData(e.pageUid);
  // Original image
  let url = URL.createObjectURL(pageData.raw.data);
  await loadImage(document.getElementById("photoRaw"), url);
  getCroppedImage();
  captureViewer.currentDocument.deleteAllPages();
});

async function loadImage(img,src) {
  return new Promise((resolve, reject) => {
    img.onload = function(){
      resolve();
    }
    img.src = src;
  })
}

Live Scan按钮被点击时启动Capture Viewer。

js 复制代码

document.getElementById("liveScanBtn").addEventListener("click",function(){
  if (captureViewer) {
    document.getElementById("captureViewer").style.display = "block";
    captureViewer.play({fill:true}); //pass fill to make set the object-fit style of the video to cover
  }
})

我们可以看到，我们定义的Handler可以和Capture Viewer一起工作了。

多边形编辑

有时我们可能需要对检测到的文档的多边形进行微调。我们可以使用Perspective Viewer来完成此操作。

为Perspective Viewer添加一个容器。
html 复制代码
```
<div id="perspectiveViewer" class="fullscreen"></div>
```

使用以下代码进行设置。

js 复制代码

let perspectiveViewer = new Dynamsoft.DDV.PerspectiveViewer({
  container: document.getElementById("perspectiveViewer"),
  uiConfig: isMobile()? mobilePerspectiveUiConfig : pcPerspectiveUiConfig
});
perspectiveViewer.openDocument(doc.uid);

它的UI（如工具栏上的按钮）可以使用UIConfig进行自定义。在这里，我们只保留perspectiveAll按钮作为确认按钮。

js 复制代码

function isMobile(){
  return "ontouchstart" in document.documentElement;
}

const mobilePerspectiveUiConfig = {
  type: Dynamsoft.DDV.Elements.Layout,
  flexDirection: "column",
  children: [
    {
      type: Dynamsoft.DDV.Elements.Layout,
      className: "ddv-perspective-viewer-header-mobile",
      children: [
        {
          // Bind event for "PerspectiveAll" button
          // The event will be registered later.
          type: Dynamsoft.DDV.Elements.PerspectiveAll,
          events:{
              click: "done"
          }
        },
      ],
    },
    Dynamsoft.DDV.Elements.MainView
  ],
};

//Pc PerspectiveViewer
const pcPerspectiveUiConfig = {
  type: Dynamsoft.DDV.Elements.Layout,
  flexDirection: "column",
  children: [
    {
      type: Dynamsoft.DDV.Elements.Layout,
      className: "ddv-perspective-viewer-header-desktop",
      children: [
        {   
          // Bind event for "PerspectiveAll" button
          // The event will be registered later.
          type: Dynamsoft.DDV.Elements.PerspectiveAll,
          className: "ddv-perspective-viewer-perspective-desktop",
          events:{
            click: "done"
          }
        },
      ],
    },
    Dynamsoft.DDV.Elements.MainView,
  ],
};

为Perspective Viewer的done事件添加侦听器。编辑完成后，使用调整过后的多边形的角点来获得裁剪的图像。

js 复制代码

perspectiveViewer.on("done", async () => {
  const quad = perspectiveViewer.getQuadSelection();
  const points = [];
  quad.forEach(p => {
    points.push({x:p[0],y:p[1]})
  });
  getCroppedImage(points);
  document.getElementById("perspectiveViewer").style.display = "none";
  perspectiveViewer.currentDocument.deleteAllPages();
});

修改Capture Viewer的captured事件的侦听器，以使用Perspective Viewer编辑多边形。

diff 复制代码

+let points = documentScanner.detect(document.getElementById("photoRaw"));
+await loadImageToDoc(document.getElementById("photoRaw").src);
+showPerspectiveViewer(points);
-getCroppedImage();
 captureViewer.currentDocument.deleteAllPages();

添加的函数：

js 复制代码

//load an image to the document and it will be displayed in the viewers
async function loadImageToDoc(dataURL) {
  const response = await fetch(dataURL);
  const blob = await response.blob();
  await doc.loadSource(blob);
}

function showPerspectiveViewer(points){
  document.getElementById("perspectiveViewer").style.display = "block";
  const quad = [];
  points.forEach(point => {
    quad.push([point.x,point.y]);
  });
  perspectiveViewer.setQuadSelection(quad);
}

屏幕截图：

打包为库

为了便于使用，我们可以将其作为库发布到NPM上。

安装devDependencies ：

bash 复制代码

npm install -D @types/node vite-plugin-dts

创建一个新的vite.config.ts文件：

ts 复制代码

// vite.config.ts
import { resolve } from 'path';
import { defineConfig } from 'vite';
import dts from 'vite-plugin-dts';
// https://vitejs.dev/guide/build.html#library-mode
export default defineConfig({
  build: {
    lib: {
      entry: resolve(__dirname, 'src/index.ts'),
      name: 'opencv-document-scanner',
      fileName: 'opencv-document-scanner',
    },
  },
  plugins: [dts()],
});

将我们的包的入口点添加到package.json。

json 复制代码

{
  "main": "./dist/opencv-document-scanner.umd.cjs",
  "module": "./dist/opencv-document-scanner.js",
  "types": "./dist/index.d.ts",
  "exports": {
    "import": {
      "types": "./dist/index.d.ts",
      "default": "./dist/opencv-document-scanner.js"
    },
    "require": {
      "types": "./dist/index.d.ts",
      "default": "./dist/opencv-document-scanner.umd.cjs"
    }
  },
  "files": [
    "dist/*.css",
    "dist/*.js",
    "dist/*.cjs",
    "dist/*.d.ts"
  ]
}

运行npm run build。然后，我们可以在dist文件夹中看到打包好的文件。

局限

该算法比较基础。在深色背景下扫描白色文档的效果很好。但背景复杂的话效果就不行了。对于更复杂的情况，我们可以使用Dynamsoft Document Normalizer来获得更好的结果。

使用OpenCV进行扫描测试（demo链接）：

使用Dynamsoft Document Normalizer进行扫描测试（demo链接）：

源代码

欢迎下载源代码并尝试使用：

github.com/tony-xlh/op...