Investigate applications-laplacian pyramid blending

Assignment Help Computer Engineering
Reference no: EM131098445

E27: Computer Vision Spring 2016 - PROJECT 2

PROJECT - LAPLACIAN PYRAMIDS AND IMAGE BLENDING

OVERVIEW

In this project, you will investigate two applications: Laplacian pyramid blending and hybrid images. You can use the former to make smooth transitions between arbitrary images, such as the fish sandwich depicted in Figure 1. The latter can be used to generate interesting optical illusions as seen in Figure 4.

1685_Figure.png

LAPLACIAN PYRAMID BLENDING

1. First, generate a Laplacian pyramid.

A Laplacian pyramid (see section 3.5.3 of the Szeliski textbook) encodes an image as a succession of progressively smaller Laplacian images, built atop a base layer consisting of a blurred and reduced copy of the original image. Given an input image I, we can construct a Laplacian pyramid P = (L0, L1, L2, . . . ,LN) of depth N + 1 according to the following rules:

  • Define G0 to be the input image itself, so G0 = I.
  • Given Gi, obtain Gi+1 by convolving Gi with a small Gaussian, and then discarding the odd-numbered rows and columns of the image, to obtain a half-sized result.
  • Given Gi+1, produce Gi+1 by enlarging Gi+1 to be the same size as Gi while smoothing the result.
  • For any i < N, define Li = Gi - Gi+1
  • Finally, let LN = GN

588_Figure1.png

When constructing the Laplacian pyramid, please keep the following tips in mind:

  • Creating Gi+1 from Gi may be accomplished via the cv2.pyrDown function.
  • Creating Gi+1 from Gi+1 may be accomplished via the cv2.pyrUp function. You will almost certainly want to supply the dstsize parameter so that Gi+1 and Gi have the same shape.
  • Be careful not to directly subtract two numpy.uint8 images - the result will certainly overflow. Instead, convert images to numpy.float32 datatype before subtracting.
  • When displaying floating-point images via cv2.imshow, OpenCV expects the intensities to vary between 0.0 and 1.0. Therefore, a reasonable method to display a floating-point Laplacian image (e.g. for debugging purposes) might be something like:

cv2.imshow(window, 0.5 + 0.5*(L / numpy.abs(L).max()))

Write a function pyr_build(img) which takes an 8-bit per channel RGB or grayscale image as input, and which outputs a list lp of Laplacian images (stored in numpy.float32 format) corresponding to the image's Laplacian pyramid. Your code should produce results similar to Figure 2.

2. Reconstruct an image from a Laplacian pyramid

Surprisingly, no information is lost when converting from an original image to the corresponding Laplacian pyramid. You can reconstruct the original image from the Laplacian pyramid (L0, L1, . . . , LN) by following these steps:

  • Let RN = LN
  • Given Ri, let Ri-1 = RI + Li-1, where RI is an enlarged and smoothed version of Ri with the same dimensions as Li-1.
  • When you reach R0, the result should be virtually indistinguishable from the original image.

Keep in mind these practical implementation matters:

  • You can use the cv2.pyrUp function to create RI from Ri. Be careful to specify destination size as you did when building the pyramid.
  • You will want to assemble R as a numpy.float32 array, but convert to numpy.uint8 immediately prior to displaying. If you have any problems with overflow, you can use the numpy.clip function to restrict R to the range [0, 255] before type-converting.

Write a function pyr_reconstruct(lp) that reconstructs the original image from a Laplacian pyramid, and test it to verify that it works before proceeding to the next section.

3. Combine two Laplacian pyramids.

Given two original images imgA and imgB, it is possible to combine their corresponding Laplacian pyramids lpA and lpB in such a way that low-frequency features (such as solid color areas) are blended over a large distance, while high-frequency features (such as fine lines, ripples or edges) are blended over much shorter distences.

One good way to combine images is with a continuously-varying mask, sometimes called an alpha mask. Let alpha be a 2D floating-point array with pixel intensities in the range [0.0, 1.0], and assume that it has the same height and width as two images A and B. Then you can use this alpha_blend function to combine the images:

def alpha_blend(A, B, alpha):

A = A.astype(alpha.dtype)

B = B.astype(alpha.dtype)

# if A and B are RGB images, we must pad

# out alpha to be the right shape

if len(A.shape) == 3:

alpha = numpy.expand_dims(alpha, 2)

return A + alpha*(B-A)

At each level of the Laplacian pyramid, resize the original alpha mask (which had the same dimensions as imgA and imgB) to the size of each of the Laplacian images using cv2.resize with interpolation mode cv2.INTER_AREA. Then, use the resized alpha mask to blend the pair of Laplacian images at that level. The result is shown in Figure 3.

Your task is to photograph (or obtain) a pair of images to blend together, and write a program to generate an alpha mask and perform Laplacian pyramid blending between the two images. One reasonable way of generating an alpha mask with OpenCV is to draw a geometric shape (i.e. using cv2.ellipse) and the smooth the resulting image using cv2.GaussianBlur. Your program should use the pyr_build and pyr_reconstruct functions you have already written.

362_Figure2.png

In addition to submitting your program and the Laplacian-pyramid-blended output, you should also submit the "traditional" result of directly alpha-blending the two input images, without the pyramid. Hopefully, the pyramid result looks much more convincing than the traditional result.

HYBRID IMAGE

Moving on from Laplacian pyramids, a related concept is the hybrid image illusion, as shown in Figure 4. A hybrid image I can be created from two images A and B according to the formula

I = kA · lopass(A, σA) + kB · hipass(B, σB)

Here, the lopass(X, σ) function implements a low-pass filter on the image X by blurring it with a Gaussian kernel with standard deviation σ, and hipass implements a high-pass filter according to

hipass(X, σ) = X - lopass(X, σ)

Typically σA > σB, and setting both kA and kB to 1 is a reasonable place to start.

Your task here is to obtain a pair of images A and B, and write a program to merge them into a hybrid image. As with the Laplacian pyramid code, you will probably want to work in numpy.float32 as an intermediate format, and be careful about overflow when converting from floating-point back to numpy.uint8 format. Please submit source code, source images, and output hybrid image.

WRITEUP

Please create a 2-4 PDF writeup addressing the following questions:

  • Who did what for this project?
  • How did you obtain and align your images for each of the two tasks? Did you use any thirdparty software (e.g. Paintbrush, Photoshop), or write a program to help prepare the images or mask?
  • What depth did you choose to build your Laplacian pyramid to, and why?
  • Why does Laplacian pyramid blending blend low-frequency content over a larger distance than high-frequency content? See if you can illustrate this with some carefully chosen input image examples.
  • How did you arrive at good values for the constants σA, σB, kA, and kB for the hybrid image generation? Describe the process.
  • If you display your hybrid image at full size on your computer screen, how close do you need to be in order to primarly see image B? How far away do you need to get before you only see features from image A? Are these distances fairly consistent between you, your lab partner, and any unsuspecting friends you show your image to?
  • What does the Laplacian pyramid of your hybrid image look like?

Please include your source images/masks, and program outputs as figures, along with whatever other images you think will be useful/informative.

EVALUATION CRITERIA

Your project will be evaluated by the following criteria (percentages are approximate):

a. source code, raw data, and output - Turn in the full source code for your project, including any additional files needed to run it. You should also submit the program outputs as image files. I expect your code to be neatly indented and reasonably commented. Please expect to spend plenty of time photographing/obtaining and aligning suitable input images for both main tasks above - it is critical to have reasonably well aligned inputs, especially for the second task!

b. Laplacian pyramid results - The results of your Laplacian pyramid blending should be free of sharp seams, and be clearly superior to traditional alpha blending.

c. hybrid image results  - Your image should clearly transition from image A to image B as the viewer approaches from far away.

d. going further - Go above and beyond the tasks outlined above. Examples might include writing a helper program to align input images and/or generate masks; adapting either one of the above techniques to work on video (live or pre-recorded); visualizing intermediate computations/results of the above tasks; researching and discussing other uses of pyramid representations in computer vision, etc. Feel free to ask if you want to run any ideas by me!

e. written report - Your PDF report should address all of the questions mentioned above, and contain the relevant images as figures.

1449_Figure3.png

Reference no: EM131098445

Detecting hidden files on suspect drive with forensics tools

What procedures are used by the suspect in order to hide the files on suspect drive, or formatted media and how does computer forensics investigator examine the hidden, pass

Identify control necessary to defend against vulnerabilities

Conduct a comprehensive and detailed analysis of techniques for hardening operating systems and applications - Describe vulnerabilities related to the failure to harden operat

What is the maximum rate

assume an 802.11b station is configured to always reserve the channel with the RTS/CTS sequence. Suppose this station suddenly wants to transmit 1,000 bytes of data, and all

Reconfiguring the network for your company

You have been tasked to reconfigure the network for your company that has grown noticeably over the past year. You have used only the static routing in past, however you do

Make a third single dimension array to hold a sum

design a third single dimension array to hold a sum. Your main program will take the two arrays of float and pass them to the function addfloat() Inside the function, add th

Why throw and handle an appropriate exception

Write down a C++ program that prompts the user to enter a length in feet and inches and outputs the equivalent length in centimeters. If the user enters a negative number or

Illustrate the architecture of the oracle instance

The Oracle Instance and Transaction/Query Processing (a) Illustrate the architecture of the Oracle instance, including all memory structures, processes, and associated physi

Conducting the forensic investigation

Describe the purpose of several initial reviews being conducted by forensics investigator (e.g., time lines, technical skills, procedures).

Reviews

Write a Review

 
Free Assignment Quote

Assured A++ Grade

Get guaranteed satisfaction & time on delivery in every assignment order you paid with us! We ensure premium quality solution document along with free turntin report!

All rights reserved! Copyrights ©2019-2020 ExpertsMind IT Educational Pvt Ltd