Image Processing in C++

-Devendra Pratap Yadav

The following image enhancements can be done using the program.

1. Rotate

2. Translate

3. Scale

4. Skew

5. Negative

6. Log Transformation

7. Inverse Log Transformation

8. Gamma Transformation

9. Contrast Stretching

10. Gray Level Slicing

11. Bit Plane Slicing

12. Histogram Equalization

13. Adaptive Histogram Equalization

14. Histogram Matching

15. Mean Filtering

16. Weighted Mean Filtering

17. Median Filtering

18. Sobel Filtering

19. Laplacian Filtering

20. HighBoost Filtering

21. Unsharp Masking

Whenever available, we compare our result to OpenCVs inbuilt function.
The program provides a menu type input to perform various functions on the image.

1. Rotate

We rotate the image using the affine transformation matrix.

Rotate by 30 degrees:

My Algorithm (Bilinear Interpolation) OpenCV

RMSE= 2.32

To rotate the image, we must use the inverse affine transform to avoid empty points in new image which dont have any point of original mapping to them.

Bilinear interpolation reduces jagginess and provides smooth image.

We conclude that bilinear interpolation is better for rotation as it provides smooth image.

2. Translate

We can translate using the affine matrix as well.

Using same original image as rotation, Translate by x=100, y=-50

Result:

The image goes out of the boundary and some data is lost while translating it.

3. Scale

Scaling can be done with affine transformation matrix as well.
We can use both nearest neighbor and bilinear interpolation.

Scale to 2x in both x and y

OpenCV Mine (bilinear)

Nearest Neighbor Original

We observe that bilinear provides smoother scaling.

4. Skew

Skewing/ Shearing can be achieved with affine matrix as well.
Output of Skew X=0.3 , Y = 0.1 using bilinear interpolation.

Bilinear Nearest Neighbor

The bilinear image has less jagged edges and provides smoother result.

5. Negative

Negative filter changes the pixel intensity as : f(x,y)= 255- f(x,y)
Result can be seen below:

ORIGINAL NEGATIVE

This can highlight interesting properties in the image.

6. Log Transformation

Logarithm transformation is : f(x,y) = c* log(1+f(x,y))
where, c= 255/log(256) to scale the values between 0 and 255.

Result can be seen below:

ORIGINAL TRANSFORMED

We observe that log transform significantly brightens the image and reveals darker details.

7. Inverse Log Transformation

It is the inverse function of the log transform function stated previously. We calculate the mapping for all gray values and then inverse the function.

Result can be seen below:

ORIGINAL TRANSFORMED

This removes the washed out, low contrast look of the image and darkens it.

8. Gamma Transformation

Logarithm transformation is :

f(x,y) = c* f(x,y)^gamma, where, c= 1 and gamma varies.
here, f(x,y) is value between 0 and 1

Result can be seen below:

ORIGINAL TRANSFORMED

For gamma=5, we get the above result. We can see that the image is significantly darkened.
For gamma<1, the image is brightened.

9. Contrast Stretching

We map the gray levels using a stretching function which takes two input points and generates the mapping as shown:

For input points (80,30) and (150,240) we get the result:

ORIGINAL TRANSFORMED

We see that the contrast is significantly improved after applying the function.

10. Gray Level Slicing

This refers to setting all pixel values to 0 except those in a given interval (low,high).

This helps to isolate image parts based on gray values

For input points low=150, high=200 we get the result:

ORIGINAL GRAY LEVEL (150, 200)

This can isolate certain gray levels of the image and help in specific analysis of image.

11. Bit Plane Slicing

Bit plane using bit k slicing refers to setting all pixels having the kth bit set to 255 and those with kth bit=0 to 0;
This separates the image into bit planes which may be combined to obtain original image.

Bit= 7 Bit = 4

This can reveal the number of pixel values above a given threshold i.e. 2^bit.

12. Histogram Equalization

Histogram equalization improves the contrast of the image using cumulative function and maps values so that pixel count is evenly distributed among all gray values.

` Original Histogram Equalized Histogram ORIGINAL Image

OPENCV MINE

RMSE= 11.5

We see that Equalization significantly improves contrast of the image

13. Adaptive Histogram Equalization

This is same as Histogram Equalization except we divide the image into blocks and compute equalized histogram for each block separately and interpolate the final result for each pixel

ORIGINAL ADAPTIVE H. E. SIMPLE H. E.

We can see that adaptive H.E. shows more details in areas where simple H.E. is very dark.
Compared with OpenCV ADE, the RMSE = 25

14. Histogram Matching

In this, we have a initial image and a final image. We have to map the histogram of the initial to the final image.
INITIAL FINAL

Matched Image

We can see that the histogram of the image is similar to final histogram and the image has better contrast than before.

15. Mean Filtering

This consists of simple spatial filtering matrix of 3x3 where each value is 1/9. This averages the value of pixel according to its neighbors.

ORIGINAL MEAN FILTERED

Comapred to OPENCV mean => RMSE= 6.31

16. Weighted Mean Filtering

Same as previous but center has weight 4, with 4 neghborhood points having weight 2 and other have weight 1.

Compared to OPENCV Gaussian Blue : RMSE = 6.1

17. Median Filtering

This sorts the points in 3x3 neighborhood and takes the median of the sorted list.

ORIGINAL MEDIAN

This reduces noise greatly as can be seen in the bottom right boxes in the image.

Comapred to OPENCV median : RMSE=8.3

18. Sobel Filtering

This uses two spatial filtering matrices to get gradients in x and y direction. Finally we get edges in the image.

ORIGINAL SOBEL

19. Laplacian Filtering

By applying the Laplacian matrix, we can sharpen the image.

ORIGINAL SHARPENED

By calculating the difference in pixel values, we can increase intensity of edges to get sharp image.

20. HighBoost Filtering

This also gives sharpened image by subtracting the blurred image from original one.

F(x,y) = f(x,y)+k*( f(x,y)-Fblur(x,y) )
The coefficient k is greater than one in this case.

ORIGINAL SHARPENED with k=5

This results in lower but better sharpening than laplacian filter.

21. Unsharp Masking

Similar to HighBoost but with k=1.

F(x,y) = f(x,y)+k*( f(x,y)-Fblur(x,y) )

ORIGINAL SHARPENED