opencv-tots

Thresholding

Segmentation technique used to seperate object from its background. Each pixel of an image is compared with a predefined threshold value

Syntax:-

_, varname = cv.threshold("<image file>",<thrshold value>,<maximum value>,<threshold techniques>)

Thresholding techniques in opencv:

• THRESH_BINARY Converts the image into 2 color representation. Pixel values greater than the threshold are white. Lesser than the threshold are black

• THRESH_BINARY_INV Inverse of the THRESH_BINARY

• THRESH_TRUNC pixel values lesser that or equal to the thershold value will not change. Values above the treshold are assigned the treshold value itself

• THRESH_TOZERO values lower than the treshold are assigned 0. Above the treshold will remain same.

*THRESH_TOZERO_INV Inverse of THRESH_TOZERO greater than treshold is 0. Lesser than threshold remoain the same

Adaptive threshold

Syntax

varname = cv2.adaptiveThreshold("<image file>",<adaptive technique>, <thresholding technique>,<blocksize>,<C value>)

: ADAPTIVE_THRESH_MEAN_C, ADAPTIVE_THRESH_GAUSSIAN_C : THRESH_BINARY, THRESH_BINARY_INV, THRESH_TRUNC, THRESH_TOZERO, THRESH_TOZERO_INV : regoin size : constant value which is used to subtract from the mean or the weighted sum

Thresholds are calculated for smaller regions instead of having an global threshold value for the entire picture. These threshold values change from region to region

Adaptive thresholding techniques in opencv:

• ADAPTIVE_THRESH_MEAN_C Threshold value is determined by the mean of a region ie blocksize X blocksize X blocksize minus C

• ADAPTIVE_THRESH_GAUSSIAN_C Is the weighted sum ie sum of( wight X pixel value) of a blocksize X blocksize region minus C

Morphological techniques

Normally performed on binary images

Requires:-

• Original image
• Kernal: matrix of fixed dimesions. Helps in deciding the nature of operation.
• mask: Converting the image to greyscale and thresholding

Techiques:-

• Dilation

dilation = cv2.dilate(mask, kernal, iterations=2)

* Dilation helps in removing noise from image 
* Increase the white area of the image (as image is binary)
* itteration is how many times u want to execute the dilation step 
* bigger the kernal better the dilation optimal (5,5) but detected boundary size increases more than it actually is
* causes overlap if the white areas are too close if the itterationa are high

• Erosion

erosion = cv2.erode(mask, kernal, iterations=1)

* Helps in reducing the baindary size of the foreground object
* Increases the size of the black around the object's boundary
* The kernal slides over the image and the value 1 is assigned only if all the values under the kernal is 1 else 0.

• MorphologyEx

  • opening

    opening = cv2.morphologyEx(mask, cv2.MORPH_OPEN, kernal)

    • Erosion -> Dilation

  • opening

    closing = cv2.morphologyEx(mask, cv2.MORPH_CLOSE, kernal)

    • Dilation -> Erosion

  • Morphological gradeint

    mg = cv2.morphologyEx(mask, cv2.MORPH_GRADIENT, kernal)

    • pixel difference between dilation and errosion

  • Morphological Tophat

    th = cv2.morphologyEx(mask, cv2.MORPH_TOPHAT, kernal)

    • Pixel difference between actual image and MORPH_OPEN

Filters

Blurs out the corners ie smoothens the image. Removes noise in the image

• Homogeneous Filters Syntax

  dst = cv2.filter2D(<image file>, <depth>, <kernel size>)

  Easiest filters. The mean off the values under the kernal matrix Each pixel value have equal weights formula = (1/k_width X k_height)[blocksize X blocksize 1's]

• Blur Syntax

  blur = cv2.blur(<image file>, kernal)

  Averaging based algorithm

• Gaussian filters

gblur = cv2.GaussianBlur(<img file>, <kernal>, <Sigmax vlaue>)

!()[https://i.stack.imgur.com/Qc4Mq.gif]
The weights are not equal.
The weights in the middel are heigher than the weights around 

• MedianBlur

  median = cv2.medianBlur(img, 5)

  This filter needs an odd numbered kernal size This filter is best for salt and pepper noise

Salt and pepper noise

Noise where in some places white noise and other places black noise !(salt)[https://www.fit.vutbr.cz/~vasicek/imagedb/img_corrupted/impnoise_005/106020.png]

Image gradients

• Laplacian

  lap = cv2.Laplacian(img, cv2.CV_64F, ksize=<kernal size>) 
  lap = np.uint8(np.absolute(lap))

  cv2.CV_64: is a datatype which allows for -ve values as well uint8 converts the image back to the unsigned int Ksize: lower the value more pronounced the boundary lines

• Soble Syntax

    sobelX = cv2.Sobel(img, cv2.CV_64F, 1, 0,ksize=kernal size)
    sobelY = cv2.Sobel(img, cv2.CV_64F, 0, 1,ksize=kernal size)
    sobelX = np.uint8(np.absolute(sobelX))
    sobelY = np.uint8(np.absolute(sobelY))

x axis = (1,0) y axis = (0,1)

The values most be converted back to unsigned int