Project 1: Image Manipulation and Computational Photography
Pranav Chopada, CS 294-26
Project Objective
The objective of the assignment is to convert the digitized Prokudin-Gorskii
glass plate images to produce a color image using image processing
techniques. This is achieved by stacking the red, green, and blue plates on
top of each other to create a color photo. The techniques used are SSD, Image
Pyramid implementation and Edge Detection Implementation.
Single Scale Implementation
The three images were separated and had to be aligned on top of each other
for best results. Out of the three image plates, blue was selected as the
benchmark as it produced the best results. Sum of Squared Differences (SSD)
technique was used to calculate the offset of red and green plates against
the blue benchmark. To align the image, SSD for all possible offsets were
calculated and the minimum SSD offset was used to transform the red and
green plate to match the blue benchmark. This approach is time consuming and
is not the optimal way to solve the problem.
Results
Red Offset: (-1,7)
Green Offset: (-1,1)
Red Offset: (1,9)
Green Offset: (0,-6)
Red Offset: (3,6)
Green Offset: (2,3)
Image Pyramid Implementation
A better approach at aligning images is by rescaling the image matrix down
by 2 until it cannot be reduced further. For every rescaled matrix, the
offset is calculated and proportionally added to the overall offset. The SSD
alignment code is reused and larger images can be processsed faster. This
increases the accuracy of offset detection as the image dimensions go up.
However, this does not work well with Emir photo where the brightness levels
are different for each of the three: red, green, and blue plates. Edge detection
is used to solve that specific problem.
Results
Church
Red Offset: (-4,58)
Green Offset: (4,25)
Three Generations
Red Offset: (12,112)
Green Offset: (15,54)
Melons
Red Offset: (14,179)
Green Offset: (12,82)
Onion Church
Red Offset: (36,108)
Green Offset: (26,52)
Train
Red Offset: (32,88)
Green Offset: (6,44)
Icon
Red Offset: (23,90)
Green Offset: (17,42)
Self Portrait
Red Offset: (-3773,176)
Green Offset: (-3780,80)
Harvesters
Red Offset: (14,124)
Green Offset: (17,60)
Sculpture
Red Offset: (-27,-3117)
Green Offset: (-11,33)
Lady
Red Offset: (12, 120)
Green Offset: (9,52)
Workshop
Red Offset: (-12,106)
Green Offset: (0,53)
Emir
Red Offset: (-1028,33)
Green Offset: (24,49)
Edge Detection Algorithm (Bells and Whistles)
Edge Detection algorithm is implemented using a gaussian blur on the entire
input image and by applying the sobel filter on the gaussian blur. Further,
the image is cropped to avoid the multiple coloration near the border of the
image. The improvement using the edge detection algorithm can be clearly seen
in the example below:
Results
Using SSD Pyramid Implementation
Edge Image of Emir
Using Edge Detection Algorithm
Cropping Photos
Photo can be cropped after alignment by 10 percent on each side to remove
the coloring problem on the borders. An example of conversion is shown below:
