Lecture 4 Flashcards
Image Filtering
techniques for modifying or enhancing an image
Types of Image Filtering
Linear filters and Non-linear filters
Which are the linear filters
image smoothing/blurring
edge detection
image sharpening
Which are the non-linear filters
morphological operators (dilation,erosion)
median filter (noise reduction)
matched filters
the value in each output pixel depends on the which values
the local values of neighboring pixels
what does a non-linear filtering
replace each pixel by a non-linear combination of its neighbors
what does linear filtering
replace each pixel by a linear combination of its neighbors
Image Filtering Steps
- move filter matrix over the image
- multiply the corresponding image pixel
- sum the results and store it as a new pixel value
Image smoothing
replace each pixel with the average intensity of its neighbors
cross-correlation
measures the similarities between two signals
convolution
measures the similarities between two signals, but the filter is flipped upside-down
which convolutions can save a lot of time
separable convolutions
strides
skip steps at a time (for efficiancy)
padding
extending the frame of an image with empty pixels, so when the kernel goes over the image we can maintain the size of the image
why we need padding
allow for more space for the kernel to cover the whole image
blurring
better image smoothing
Filters used for better image filters
gaussian filters
what are doing the gaussian filters
remove the high-spatial frequencies
edge detection
edges are sudden transition of intensity and are very informative
laplacian: measures the 2nd derivative
sensitive to noise
2 filters => sober and robert filters
subtract two blurred versions of the same image
Laplacian of Gaussian-filtered image (LoG)
measure of the 2nd derivative (the change) of the image
What is shown by the 2nd derivative in Laplacian?
showing the magnitude of a change
What does the output of Laplacian stands for?
large and positive on one side of the edge
large and negative on the other
small/zero at the edge itself
image sharpening
emphasize details of images:
edges and local contrast
what is image sharpening
emphasize the details of an image
Image + alpha*EdgeDetection
dilation
takes the max value of all pixels in its neighborhood
erosion
takes the min value of all pixels in its neighborhood
non-linear morphological operators
dilation
erosion
what does the median filter
noise reduction
median filter
replace each pixel with the median value within its neighbor
common models for noise reduction
gaussian additive noise
salt-pepper noise
Mexican Hat Filter
indicates the shape obtained by subtracting two Gaussians (“Difference of Gaussians”) with different standard deviations.
Because “Mexican hat” refers to the shape it is sometimes also used when referring to the shape of the Laplacian filter which is very similar to that of the DoG filter.
