關於邊緣檢測的基礎來自於一個事實，即在邊緣部分，畫素值出現”跳躍“或者較大的變化。如果在此邊緣部分求取一階導數，就會看到極值的出現。
而在一階導數為極值的地方，二階導數為0，基於這個原理，就可以進行邊緣檢測。
關於 Laplace 演算法原理，可參考
0x01. Laplace 演算法
下面的程式碼展示了分別對灰度化的影象和原始彩色影象中的邊緣進行檢測：
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 |
import cv2.cv as cv im=cv.LoadImage('img/building.png', cv.CV_LOAD_IMAGE_COLOR) # Laplace on a gray scale picture gray = cv.CreateImage(cv.GetSize(im), 8, 1) cv.CvtColor(im, gray, cv.CV_BGR2GRAY) aperture=3 dst = cv.CreateImage(cv.GetSize(gray), cv.IPL_DEPTH_32F, 1) cv.Laplace(gray, dst,aperture) cv.Convert(dst,gray) thresholded = cv.CloneImage(im) cv.Threshold(im, thresholded, 50, 255, cv.CV_THRESH_BINARY_INV) cv.ShowImage('Laplaced grayscale',gray) #------------------------------------ # Laplace on color planes = [cv.CreateImage(cv.GetSize(im), 8, 1) for i in range(3)] laplace = cv.CreateImage(cv.GetSize(im), cv.IPL_DEPTH_16S, 1) colorlaplace = cv.CreateImage(cv.GetSize(im), 8, 3) cv.Split(im, planes[0], planes[1], planes[2], None) #Split channels to apply laplace on each for plane in planes: cv.Laplace(plane, laplace, 3) cv.ConvertScaleAbs(laplace, plane, 1, 0) cv.Merge(planes[0], planes[1], planes[2], None, colorlaplace) cv.ShowImage('Laplace Color', colorlaplace) #------------------------------------- cv.WaitKey(0) |
效果展示
原圖
灰度化圖片檢測
原始彩色圖片檢測
0x02. Sobel 演算法
Sobel 也是很常用的一種輪廓識別的演算法。
關於 Sobel 導數原理的介紹，可參考
以下是使用 Sobel 演算法進行輪廓檢測的程式碼和效果
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 |
import cv2.cv as cv im=cv.LoadImage('img/building.png', cv.CV_LOAD_IMAGE_GRAYSCALE) sobx = cv.CreateImage(cv.GetSize(im), cv.IPL_DEPTH_16S, 1) cv.Sobel(im, sobx, 1, 0, 3) #Sobel with x-order=1 soby = cv.CreateImage(cv.GetSize(im), cv.IPL_DEPTH_16S, 1) cv.Sobel(im, soby, 0, 1, 3) #Sobel withy-oder=1 cv.Abs(sobx, sobx) cv.Abs(soby, soby) result = cv.CloneImage(im) cv.Add(sobx, soby, result) #Add the two results together. cv.Threshold(result, result, 100, 255, cv.CV_THRESH_BINARY_INV) cv.ShowImage('Image', im) cv.ShowImage('Result', result) cv.WaitKey(0) |
處理之後效果圖（感覺比Laplace效果要好些）
0x03. cv.MorphologyEx
cv.MorphologyEx 是另外一種邊緣檢測的演算法
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 |
import cv2.cv as cv image=cv.LoadImage('img/build.png', cv.CV_LOAD_IMAGE_GRAYSCALE) #Get edges morphed = cv.CloneImage(image) cv.MorphologyEx(image, morphed, None, None, cv.CV_MOP_GRADIENT) # Apply a dilate - Erode cv.Threshold(morphed, morphed, 30, 255, cv.CV_THRESH_BINARY_INV) cv.ShowImage("Image", image) cv.ShowImage("Morphed", morphed) cv.WaitKey(0) |
0x04. Canny 邊緣檢測
Canny 演算法可以對直線邊界做出很好的檢測；
關於 Canny 演算法原理的描述，可參考：
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 |
import cv2.cv as cv import math im=cv.LoadImage('img/road.png', cv.CV_LOAD_IMAGE_GRAYSCALE) pi = math.pi #Pi value dst = cv.CreateImage(cv.GetSize(im), 8, 1) cv.Canny(im, dst, 200, 200) cv.Threshold(dst, dst, 100, 255, cv.CV_THRESH_BINARY) #---- Standard ---- color_dst_standard = cv.CreateImage(cv.GetSize(im), 8, 3) cv.CvtColor(im, color_dst_standard, cv.CV_GRAY2BGR)#Create output image in RGB to put red lines lines = cv.HoughLines2(dst, cv.CreateMemStorage(0), cv.CV_HOUGH_STANDARD, 1, pi / 180, 100, 0, 0) for (rho, theta) in lines[:100]: a = math.cos(theta) #Calculate orientation in order to print them b = math.sin(theta) x0 = a * rho y0 = b * rho pt1 = (cv.Round(x0 + 1000*(-b)), cv.Round(y0 + 1000*(a))) pt2 = (cv.Round(x0 - 1000*(-b)), cv.Round(y0 - 1000*(a))) cv.Line(color_dst_standard, pt1, pt2, cv.CV_RGB(255, 0, 0), 2, 4) #Draw the line #---- Probabilistic ---- color_dst_proba = cv.CreateImage(cv.GetSize(im), 8, 3) cv.CvtColor(im, color_dst_proba, cv.CV_GRAY2BGR) # idem rho=1 theta=pi/180 thresh = 50 minLength= 120 # Values can be changed approximately to fit your image edges maxGap= 20 lines = cv.HoughLines2(dst, cv.CreateMemStorage(0), cv.CV_HOUGH_PROBABILISTIC, rho, theta, thresh, minLength, maxGap) for line in lines: cv.Line(color_dst_proba, line[0], line[1], cv.CV_RGB(255, 0, 0), 2, 8) cv.ShowImage('Image',im) cv.ShowImage("Cannied", dst) cv.ShowImage("Hough Standard", color_dst_standard) cv.ShowImage("Hough Probabilistic", color_dst_proba) cv.WaitKey(0) |
原圖
使用 Canny 演算法處理之後
標記出標準的直線
標記出所有可能的直線
0x05. 輪廓檢測
OpenCV 提供一個 FindContours 函式可以用來檢測出影象中物件的輪廓：
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 |
import cv2.cv as cv orig = cv.LoadImage('img/build.png', cv.CV_LOAD_IMAGE_COLOR) im = cv.CreateImage(cv.GetSize(orig), 8, 1) cv.CvtColor(orig, im, cv.CV_BGR2GRAY) #Keep the original in colour to draw contours in the end cv.Threshold(im, im, 128, 255, cv.CV_THRESH_BINARY) cv.ShowImage("Threshold 1", im) element = cv.CreateStructuringElementEx(5*2+1, 5*2+1, 5, 5, cv.CV_SHAPE_RECT) cv.MorphologyEx(im, im, None, element, cv.CV_MOP_OPEN) #Open and close to make appear contours cv.MorphologyEx(im, im, None, element, cv.CV_MOP_CLOSE) cv.Threshold(im, im, 128, 255, cv.CV_THRESH_BINARY_INV) cv.ShowImage("After MorphologyEx", im) # -------------------------------- vals = cv.CloneImage(im) #Make a clone because FindContours can modify the image contours=cv.FindContours(vals, cv.CreateMemStorage(0), cv.CV_RETR_LIST, cv.CV_CHAIN_APPROX_SIMPLE, (0,0)) _red = (0, 0, 255); #Red for external contours _green = (0, 255, 0);# Gren internal contours levels=2 #1 contours drawn, 2 internal contours as well, 3 ... cv.DrawContours (orig, contours, _red, _green, levels, 2, cv.CV_FILLED) #Draw contours on the colour image cv.ShowImage("Image", orig) cv.WaitKey(0) |
效果圖：
原圖
識別結果
0x06. 邊界檢測
|
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 |
import cv2.cv as cv im = cv.LoadImage("img/build.png", cv.CV_LOAD_IMAGE_GRAYSCALE) dst_32f = cv.CreateImage(cv.GetSize(im), cv.IPL_DEPTH_32F, 1) neighbourhood = 3 aperture = 3 k = 0.01 maxStrength = 0.0 threshold = 0.01 nonMaxSize = 3 cv.CornerHarris(im, dst_32f, neighbourhood, aperture, k) minv, maxv, minl, maxl = cv.MinMaxLoc(dst_32f) dilated = cv.CloneImage(dst_32f) cv.Dilate(dst_32f, dilated) # By this way we are sure that pixel with local max value will not be changed, and all the others will localMax = cv.CreateMat(dst_32f.height, dst_32f.width, cv.CV_8U) cv.Cmp(dst_32f, dilated, localMax, cv.CV_CMP_EQ) #compare allow to keep only non modified pixel which are local maximum values which are corners. threshold = 0.01 * maxv cv.Threshold(dst_32f, dst_32f, threshold, 255, cv.CV_THRESH_BINARY) cornerMap = cv.CreateMat(dst_32f.height, dst_32f.width, cv.CV_8U) cv.Convert(dst_32f, cornerMap) #Convert to make the and cv.And(cornerMap, localMax, cornerMap) #Delete all modified pixels radius = 3 thickness = 2 l = [] for x in range(cornerMap.height): #Create the list of point take all pixel that are not 0 (so not black) for y in range(cornerMap.width): if cornerMap[x,y]: l.append((y,x)) for center in l: cv.Circle(im, center, radius, (255,255,255), thickness) cv.ShowImage("Image", im) cv.ShowImage("CornerHarris Result", dst_32f) cv.ShowImage("Unique Points after Dilatation/CMP/And", cornerMap) cv.WaitKey(0) |