我在 Canvas 上展示了许多医学图片,如下所示。
我正在尝试制作一个工具,允许您使用该工具以可扩展圆圈的形式选择图像的任何区域,并仅填充不超出轮廓的部分原始点击像素所在的位置。在单独的 Canvas 层上绘制填充轮廓。
现在我使用最常见的具有可变容差(比较函数)的洪水填充迭代堆栈实现。 You can familiarize yourself with it here 。一切都不是很好,特别是在没有强烈对比度差异的图片和高分辨率图像中,其他一切都非常慢。
我有想法创建一个状态容器并查找那里是否存在所需的填充轮廓,如果存在,则只需替换 Canvas 像素数组(不过,我将不得不再次求助为了进行一些额外的处理, Canvas 像素数组包含 4 个 channel ,而在算法的输出中仅获得 1 个 channel ,仅替换内容是行不通的,您需要将每个像素替换为分为 4 个 channel 的像素)而不是每次都会缓慢地充满洪水。但这种方法有一个重大问题:内存消耗。正如您可能猜到的那样,填充的轮廓,尤其是仅具有不错的分辨率的轮廓,可能会占用相当多的空间,并且它们的设置成为内存消耗的真正问题。
决定以多边形的形式存储完成的轮廓,并从容器中提取它们,只需用更快的 context fill 填充它们即可。 。使用的算法允许我输出一组边界,但由于算法的特点,这个数组是无序的,并且按这个顺序连接顶点,我们只得到部分填充的轮廓(右图)。有没有一种方法可以对它们进行排序,使我只能连接它们并获得闭合路径(左图中填充轮廓中的孔不应该是先验的,所以我们不必担心关于他们)?
总结一下,由于填充工作不太好,我想使用不同的算法/实现,但我不知道是哪一个。 以下是我的一些想法:
使用不同的实现方式,例如行扫描方法。据我所知,here is one of the fastest and most effective implementations of the algorithm among open sources 。 优点:可能的效率和速度。 缺点:我需要以某种方式将结果转换为多边形,将算法重写为javascript(可能是 emscripten ,可以做得很好,但无论如何我都必须重写相当一部分代码)。
使用完全不同的方法。
a) 我不知道,但也许 Canny 检测器对于提取多边形很有用。但就程序在客户端的使用而言,提取所有边界是无利可图的,有必要弄清楚如何只处理必要的部分,而不是整个图片。
b) 然后,了解边界后,使用任何足够快的填充算法,该算法根本不会超出找到的边界。
我很高兴了解其他一些方法,甚至更好地了解 JavaScript 中现成的实现
更新:
为了更好地理解,工具光标和算法的预期结果如下所示。
最佳答案
这是一个使用 opencv 的示例
下面应该可以工作或最终使用代码片段内提供的 fiddle 链接
感兴趣:approxPolyDP,它可能足以满足您的需求(检查Ramer-Douglas-Peucker algorithm)
// USE FIDDLE
// https://jsfiddle.net/c7xrq1uy/
async function loadSomeImage() {
const ctx = document.querySelector('#imageSrc').getContext('2d')
ctx.fillStyle = 'black'
const img = new Image()
img.crossOrigin = ''
img.src = 'https://cors-anywhere.herokuapp.com//image/aiZ7z.png'
img.onload = () => {
const imgwidth = img.offsetWidth
const imgheight = img.offsetHeight
ctx.drawImage(img, 0, 0, img.width, img.height, 0, 0, 400, 400)
}
}
function plotPoints(canvas, points, color = 'green', hold = false){
const ctx = canvas.getContext('2d')
!hold && ctx.clearRect(0, 0, 400, 400)
ctx.strokeStyle = color
Object.values(points).forEach(ps => {
ctx.beginPath()
ctx.moveTo(ps[0].x, ps[0].y)
ps.slice(1).forEach(({ x, y }) => ctx.lineTo(x,y))
ctx.closePath()
ctx.stroke()
})
}
const binarize = (src, threshold) => {
cv.cvtColor(src, src, cv.COLOR_RGB2GRAY, 0)
const dst = new cv.Mat()
src.convertTo(dst, cv.CV_8U)
cv.threshold(src, dst, threshold, 255, cv.THRESH_BINARY_INV)
cv.imshow('binary', dst)
return dst
}
const flip = src => {
const dst = new cv.Mat()
cv.threshold(src, dst, 128, 255, cv.THRESH_BINARY_INV)
cv.imshow('flip', dst)
return dst
}
const dilate = (src) => {
const dst = new cv.Mat()
let M = cv.Mat.ones(3, 3, cv.CV_8U)
let anchor = new cv.Point(-1, -1)
cv.dilate(src, dst, M, anchor, 1, cv.BORDER_CONSTANT, cv.morphologyDefaultBorderValue())
M.delete()
cv.imshow('dilate', dst)
return dst
}
const PARAMS = {
threshold: 102,
anchor: { x: 180, y: 180 },
eps: 1e-2
}
const dumpParams = ({ threshold, anchor, eps }) => {
document.querySelector('#params').innerHTML = `thres=${threshold} (x,y)=(${anchor.x}, ${anchor.y}) eps:${eps}`
}
document.querySelector('input[type=range]').onmouseup = e => {
PARAMS.threshold = Math.round(parseInt(e.target.value, 10) / 100 * 255)
dumpParams(PARAMS)
runCv(PARAMS)
}
document.querySelector('input[type=value]').onchange = e => {
PARAMS.eps = parseFloat(e.target.value)
dumpParams(PARAMS)
runCv(PARAMS)
}
document.querySelector('#imageSrc').onclick = e => {
const rect = e.target.getBoundingClientRect()
PARAMS.anchor = {
x: e.clientX - rect.left,
y: e.clientY - rect.top
}
dumpParams(PARAMS)
runCv(PARAMS)
}
const contourToPoints = cnt => {
const arr = []
for (let j = 0; j < cnt.data32S.length; j += 2){
let p = {}
p.x = cnt.data32S[j]
p.y = cnt.data32S[j+1]
arr.push(p)
}
return arr
}
loadSomeImage()
dumpParams(PARAMS)
let CVREADY
const cvReady = new Promise((resolve, reject) => CVREADY = resolve)
const runCv = async ({ threshold, anchor, eps }) => {
await cvReady
const canvasFinal = document.querySelector('#final')
const mat = cv.imread(document.querySelector('#imageSrc'))
const binaryImg = binarize(mat, threshold, 'binary')
const blurredImg = dilate(binaryImg)
const flipImg = flip(blurredImg)
var contours = new cv.MatVector()
const hierarchy = new cv.Mat
cv.findContours(flipImg, contours, hierarchy, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
const points = {}
let matchingPoints = null
let matchingContour = null
for (let i = 0; i < contours.size(); ++i) {
let minArea = 1e40
const ci = contours.get(i)
points[i] = contourToPoints(ci)
if (anchor) {
const point = new cv.Point(anchor.x, anchor.y)
const inside = cv.pointPolygonTest(ci, point, false) >= 1
const area = cv.contourArea(ci)
if (inside && area < minArea) {
matchingPoints = points[i]
matchingContour = ci
minArea = area
}
}
}
plotPoints(canvasFinal, points)
if (anchor) {
if (matchingPoints) {
plotPoints(canvasFinal, [matchingPoints], 'red', true)
if (eps) {
const epsilon = eps * cv.arcLength(matchingContour, true)
const approx = new cv.Mat()
cv.approxPolyDP(matchingContour, approx, epsilon, true)
const arr = contourToPoints(approx)
console.log('polygon', arr)
plotPoints(canvasFinal, [arr], 'blue', true)
}
}
}
mat.delete()
contours.delete()
hierarchy.delete()
binaryImg.delete()
blurredImg.delete()
flipImg.delete()
}
function onOpenCvReady() {
cv['onRuntimeInitialized'] = () => {console.log('cvready'); CVREADY(); runCv(PARAMS)}
}
// just so we can load async script
var script = document.createElement('script');
script.onload = onOpenCvReady
script.src = 'https://docs.opencv.org/master/opencv.js';
document.head.appendChild(script)
canvas{border: 1px solid black;}
.debug{width: 200px; height: 200px;}
binarization threeshold<input type="range" min="0" max="100"/><br/>
eps(approxPolyDp) <input type="value" placeholder="0.01"/><br/>
params: <span id="params"></span><br/>
<br/>
<canvas id="imageSrc" height="400" width="400"/></canvas>
<canvas id="final" height="400" width="400"></canvas>
<br/>
<canvas class="debug" id="binary" height="400" width="400" title="binary"></canvas>
<canvas class="debug" id="dilate" height="400" width="400" title="dilate"></canvas>
<canvas class="debug" id="flip" height="400" width="400" title="flip"></canvas>
ps:控制台输出多边形
使用掩码实现
编辑:在下面的代码片段中,我获得了更多乐趣并实现了蒙版。我们可以将片段制作为[整页],然后将鼠标悬停在第一个 Canvas 上。
// USE FIDDLE
// https://jsfiddle.net/c7xrq1uy/
async function loadSomeImage() {
const ctx = document.querySelector('#imageSrc').getContext('2d')
ctx.fillStyle = 'black'
const img = new Image()
img.crossOrigin = ''
img.src = 'https://cors-anywhere.herokuapp.com//image/aiZ7z.png'
img.onload = () => {
const imgwidth = img.offsetWidth
const imgheight = img.offsetHeight
ctx.drawImage(img, 0, 0, img.width, img.height, 0, 0, 400, 400)
}
}
function plotPoints(canvas, points, color = 'green', hold = false){
const ctx = canvas.getContext('2d')
!hold && ctx.clearRect(0, 0, 400, 400)
ctx.strokeStyle = color
Object.values(points).forEach(ps => {
ctx.beginPath()
ctx.moveTo(ps[0].x, ps[0].y)
ps.slice(1).forEach(({ x, y }) => ctx.lineTo(x,y))
ctx.closePath()
ctx.stroke()
})
}
const binarize = (src, threshold) => {
cv.cvtColor(src, src, cv.COLOR_RGB2GRAY, 0)
const dst = new cv.Mat()
src.convertTo(dst, cv.CV_8U)
cv.threshold(src, dst, threshold, 255, cv.THRESH_BINARY_INV)
cv.imshow('binary', dst)
return dst
}
const flip = src => {
const dst = new cv.Mat()
cv.threshold(src, dst, 128, 255, cv.THRESH_BINARY_INV)
cv.imshow('flip', dst)
return dst
}
const dilate = (src) => {
const dst = new cv.Mat()
let M = cv.Mat.ones(3, 3, cv.CV_8U)
let anchor = new cv.Point(-1, -1)
cv.dilate(src, dst, M, anchor, 1, cv.BORDER_CONSTANT, cv.morphologyDefaultBorderValue())
M.delete()
cv.imshow('dilate', dst)
return dst
}
const PARAMS = {
threshold: 102,
anchor: { x: 180, y: 180 },
eps: 1e-2,
radius: 50
}
const dumpParams = ({ threshold, anchor, eps }) => {
document.querySelector('#params').innerHTML = `thres=${threshold} (x,y)=(${anchor.x}, ${anchor.y}) eps:${eps}`
}
document.querySelector('input[type=range]').onmouseup = e => {
PARAMS.threshold = Math.round(parseInt(e.target.value, 10) / 100 * 255)
dumpParams(PARAMS)
runCv(PARAMS)
}
document.querySelector('input[type=value]').onchange = e => {
PARAMS.eps = parseFloat(e.target.value)
dumpParams(PARAMS)
runCv(PARAMS)
}
document.querySelector('#imageSrc').onclick = e => {
const rect = e.target.getBoundingClientRect()
PARAMS.anchor = {
x: e.clientX - rect.left,
y: e.clientY - rect.top
}
dumpParams(PARAMS)
runCv(PARAMS)
}
// sorry for the globals, keep code simple
let DST = null
let MATCHING_CONTOUR = null
let DEBOUNCE = 0
document.querySelector('#imageSrc').onmousemove = e => {
if (Date.now() - DEBOUNCE < 100) return
if (!MATCHING_CONTOUR || !DST) { return }
const rect = e.target.getBoundingClientRect()
DEBOUNCE = Date.now()
const x = e.clientX - rect.left
const y = e.clientY - rect.top
const dst = DST.clone()
plotIntersectingMask(dst, MATCHING_CONTOUR, { anchor: { x, y }, radius: PARAMS.radius })
dst.delete()
}
const contourToPoints = cnt => {
const arr = []
for (let j = 0; j < cnt.data32S.length; j += 2){
let p = {}
p.x = cnt.data32S[j]
p.y = cnt.data32S[j+1]
arr.push(p)
}
return arr
}
const plotIntersectingMask = (dst, cnt, { anchor, radius }) => {
const { width, height } = dst.size()
const contourMask = new cv.Mat.zeros(height, width, dst.type())
const matVec = new cv.MatVector()
matVec.push_back(cnt)
cv.fillPoly(contourMask, matVec, [255, 255, 255, 255])
const userCircle = new cv.Mat.zeros(height, width, dst.type())
cv.circle(userCircle, new cv.Point(anchor.x, anchor.y), radius, [255, 128, 68, 255], -2)
const commonMask = new cv.Mat.zeros(height, width, dst.type())
cv.bitwise_and(contourMask, userCircle, commonMask)
userCircle.copyTo(dst, commonMask)
cv.imshow('final', dst)
commonMask.delete()
matVec.delete()
contourMask.delete()
userCircle.delete()
}
loadSomeImage()
dumpParams(PARAMS)
let CVREADY
const cvReady = new Promise((resolve, reject) => CVREADY = resolve)
const runCv = async ({ threshold, anchor, eps, radius }) => {
await cvReady
const canvasFinal = document.querySelector('#final')
const mat = cv.imread(document.querySelector('#imageSrc'))
const binaryImg = binarize(mat, threshold, 'binary')
const blurredImg = dilate(binaryImg)
const flipImg = flip(blurredImg)
var contours = new cv.MatVector()
const hierarchy = new cv.Mat
cv.findContours(flipImg, contours, hierarchy, cv.RETR_EXTERNAL, cv.CHAIN_APPROX_SIMPLE)
const points = {}
let matchingPoints = null
let matchingContour = null
for (let i = 0; i < contours.size(); ++i) {
let minArea = 1e40
const ci = contours.get(i)
points[i] = contourToPoints(ci)
if (anchor) {
const point = new cv.Point(anchor.x, anchor.y)
const inside = cv.pointPolygonTest(ci, point, false) >= 1
const area = cv.contourArea(ci)
if (inside && area < minArea) {
matchingPoints = points[i]
matchingContour = ci
minArea = area
}
}
}
plotPoints(canvasFinal, points)
if (anchor) {
if (matchingPoints) {
MATCHING_CONTOUR = matchingContour
plotPoints(canvasFinal, [matchingPoints], 'red', true)
if (eps) {
const epsilon = eps * cv.arcLength(matchingContour, true)
const approx = new cv.Mat()
cv.approxPolyDP(matchingContour, approx, epsilon, true)
const arr = contourToPoints(approx)
//console.log('polygon', arr)
plotPoints(canvasFinal, [arr], 'blue', true)
if (DST) DST.delete()
DST = cv.imread(document.querySelector('#final'))
}
}
}
mat.delete()
contours.delete()
hierarchy.delete()
binaryImg.delete()
blurredImg.delete()
flipImg.delete()
}
function onOpenCvReady() {
cv['onRuntimeInitialized'] = () => {console.log('cvready'); CVREADY(); runCv(PARAMS)}
}
// just so we can load async script
var script = document.createElement('script');
script.onload = onOpenCvReady
script.src = 'https://docs.opencv.org/master/opencv.js';
document.head.appendChild(script)
canvas{border: 1px solid black;}
.debug{width: 200px; height: 200px;}
#imageSrc{cursor: pointer;}
binarization threeshold<input type="range" min="0" max="100"/><br/>
eps(approxPolyDp) <input type="value" placeholder="0.01"/><br/>
params: <span id="params"></span><br/>
<br/>
<canvas id="imageSrc" height="400" width="400"/></canvas>
<canvas id="final" height="400" width="400"></canvas>
<br/>
<canvas class="debug" id="binary" height="400" width="400" title="binary"></canvas>
<canvas class="debug" id="dilate" height="400" width="400" title="dilate"></canvas>
<canvas class="debug" id="flip" height="400" width="400" title="flip"></canvas>
关于javascript - 一种快速填充轮廓的方法,并且能够导出为类似多边形的格式,我们在Stack Overflow上找到一个类似的问题: https://stackoverflow.com/questions/61685407/