1. Image enhancement Last update 2015.03.22 Heejune Ahn, SeoulTech

2. Outline Purpose of image enhancment Target Pixel, neighbor hood, filter Linear Filtering Noise Reduction Filters Edge Detection Filters Edge Enhancement Filters

3. 1. Image enhancement Purpose  clarify the information of image  depends upon what to extract Methods  (spatial-domain) filtering: linear or non linear  Freq-domain processing (in Ch5) Info 1 Info 2 Info 3.. Noise Info 1 Info 2 Info 3.. Noise Enha ncem ent original image enhanced image By Human Or Machine Info 1

4. 2. Target, Neighborhood, Filter Target pixel  the pixel for process/interest Connectivity  4 connectivity: N, W, E, S  8 connectivity NB (Neighborhood)  region of (i +/-k, j+/-k), 0<= k <= floor(N/2) NB operation  function(NB of (x,y))  nfilter(Image, range, filterfunc) In MATLAB  Note: function handle: e.g. func = @(x) max(x[:])

5. 3. Linear Filter and Filter Kernel Step  align center of mask and target pixel  Calculate the weighted sum  Assign it to target pixel mask (kernel) = matrix of weights center of mask target pixel

6. In convolution notation or Adaptive filtering  Kernel changes depending on the environments

7. Image boundary  No pixels in NB  Solutions Leave unchanged  Clip image or not Use only available pixels Fill in missing pixels (e.g. replicates) Nonlinear spatial filter  Can in convolution expr.  Cannot in convolution expr. E.g.) ordering, sorting, etc

8. 4. Filter for Noise Reduction Noise type  Salt-pepper, Gaussian Noise simulation  For testing, since we cannot real noise easily.  imnoise(I, type, param) in MATLAB E.g. (‘salt & pepper’, %), (‘gaussian’, variance)

9. Mean-filter  Kernel:  Characteristics Smoothing (low pass), maintaining same mean value Kernel size : as large, blurring (loss of details) Noise removal perform  good for Gaussian noise, not good for S&P noise Median-Filter  Kernel = median (NB)  Characteristics Preserve sharp edge, remove isolated values

10.  Noise removal: ok for Gaussian, good for S&P.  Heavy computation: ordering/sorting at every pixel Rank-filter  Generalized median filter  Kernel = rank_n(NB of pixel(x,y))  Special type: max, min filter  Conservative filter I(i) p(i) I(i) o.w.

11. Gaussian Filter  Kernel  STD: degree of smoothing  Similar to mean filter (LPF) Noise reduction before edge detection  Mathematical benefit Gaussian func. Gaussian func. F

12. 5. Edge Detection Filter Derivate-filters for discontinuity detection  cf) mean/smoothing filters are integration filters   Types : 1 st order, 2 nd order different objects Discontinuity of pixel values in boundary edges

13.  Complete smooth region: 0 Computational characteristics  Linear operators: kernel implementation  LoG filter Laplacian with Gaussian filter, very popular in feature extraction (e.g SIFT) Not ‘log’arithmic

14. First order Edgedetector

15. Computational properties  Separable Filters 2D Fillter O(N 2 ) 1D Fillter O(N) 1D Fillter O(N) O(2N)

16. 2 nd order Edge Detector: Laplacian  detect of change of gradient Boundary at zero crossing Sharp edge than 1 st order detector

17. Comparisons of edge detectors

18. Kernel Reduction of Noise sensitivity  Integrated with Gaussian filter : LoG In MATLAB  edge(Img, ‘zerocross’, [], filer) Gaussian (STD) Laplacian LoG

19. 6. Edge enhancement Visual enhancement of boundary  For human perception Methods  Laplacian edge sharpening  Unsharp mask

20. example