E9241-Assignment 02 Solved

Q1. Contrast Enhancement: For the images IIScMainBuilding_LowContrast.png, LowLight_2. png, LowLight_3.png, Hazy.png, StoneFace.png, apply the below contrast enhancement algorithms. For every image, mention which of the contrast enhancement algorithms succeeded and which of them failed. Also, explain why the algorithm succeeded or failed.

(a)    Full Scale Contrast Stretching.

(b)    Non-linear Contrast Stretching: Use logarithm and exponential operators (apply log or expon each pixel intensity). Make sure the output image intensities are in the valid range.

(c)    Histogram Equalization.

(d)   Contrast Limited Adaptive Histogram Equalization (CLAHE): Divide the image into 8 × 8 non overlapping blocks (for eg, for an 80×80 image, there are 64 blocks of size 10×10. There will always be 64 blocks only). Perform Histogram equalization on each block. Also divide the image into blocks with 25% overlap. Compare both the algorithms.

For all the five algorithms, plot the image and the histogram before and after enhancement. Function inputs: grayscale image

Function outputs: contrast enhanced images and the corresponding histograms.

          .                                                                                                                                                      (5+10+10+20=45M)

Q2. Image Upsampling: Write a function to upsample a grayscale image by a factor of k using either nearest neighbor interpolation or bilinear interpolation. Which method gives better results? One way to quantify this is by first subsampling a given image I by a factor of k. Then upsample the subsampled image by the same factor k by using both interpolation methods. Now, compute mean squared error (MSE) between the original image and the upsampled image and compare the errors.

Perform this operation for images StoneFace.png and Bee.jpg for scaling factors k = 2,3 using both interpolation methods. Plot the original and upsampled images and also provide the MSE values.

Function inputs: subsampled grayscale image, scaling factor (k) and interpolation method

            Function outputs: upsampled image.                                                                                                           

1

Q3. Image Rotation: Write a function to rotate the image Bee.jpg in the clockwise direction by 30◦, 75◦ and in the counter-clockwise direction by 45◦. Make sure the rotated image is completely visible and not cropped.

Function inputs: grayscale image, angle of rotation (in degrees, positive for clockwise and negative for counter-clockwise) and interpolation method

             Function outputs: rotated image.                                                                                                                 

Q4. Spatial Domain Filtering:

(a)    Mitigate the noise in the image noisy.tif by filtering with a square averaging mask of sizes 5, 10, 15. What do you notice with increasing mask size?

(b)    Use high boost filtering to sharpen the denoised image from part (a). Choose the scalingconstant for the high pass component that minimizes the mean squared error between the sharpened image and the clean image characters.tif.

(c)    What happens if you apply high boost filtering directly on noisy.tif without smoothing?

Function inputs: noisy image Function outputs: smooothed and sharpened images.      

Q5. Histogram Matching for Object Detection (optional bonus): There has been a crime in a remote village and the offender has escaped the village in a car around sunset. The police did not find any clues in the village that could lead to the suspect except for a picture. While escaping, the suspect’s car has been caught on a security camera installed near the road (crime_original.png). However, there are no other cameras for another 200 kms. The police blocked all exit routes by night, but probably the suspect has escaped by that time. Unable to solve this crime, Inspector Lestrade approached the only consulting detective available, Sherlock Holmes!

To figure out the direction in which the suspect fled, Sherlock collected images from all security cameras in the roads leading out of the village. But there are so many images and it is time consuming to check each image one by one. As you know, Sherlock is busy with other cases and nevertheless, Sherlock isn’t interested in such mundane tasks. Sherlock has heard that there are some digital image processing (DIP) techniques which can identify the car automatically. Obviously, Sherlock knows nothing about image processing and has put up an advertisement for the development of an algorithm that can automatically detect the car. A DIP rookie that Lestrade knows tried to solve this using template matching (explained below). However, because the images have different lighting conditions, template matching is giving garbage results. Since you’re a DIP expert, Sherlock and Lestrade are hopeful that you can come up with something better. Can you help Sherlock in this task?

Since Sherlock cannot provide all the original images, he has put up a few sample images crime_ sample1.png, crime_sample2.png, crime_sample3.png and crime_sample4.png. To make job easier for you, the car image from crime_original.png has been extracted and provided for you (crime_car.png). Write a function to automatically detect the car in the sample images and put a bounding box around the car in the image (use cv2.rectangle() function). Hint: Use histogram matching before applying template matching.

Display the original image (crime_original.png) and the automatically detected image with a bounding box over the car for both vanilla template matching and for template matching with histogram matching. Also display the four sample images after histogram matching. Explain why template matching fails in this case and how preprocessing the image with histogram matching helps. (You need to implement both template matching and histogram matching yourself).

Function inputs: the images crime_original.png, crime_car.png, crime_sample1.png, crime_ sample2.png, crime_sample3.png and crime_sample4.png

Function outputs: The images detected by vanilla template matching and histogram matching followed by template matching and the histogram matched sample images.

Template Matching: Given an image I and a pattern P, extract all possible windows W of size same as P from the image I and compute the mean squared error (MSE) between the pattern P and windows W. Whichever window W has least MSE with the pattern P is your output i.e. this window contains a pattern similar to the one in P. 

2