CS 143 / Project 1 / Image Filtering and Hybrid Images
Introduction
Hybrid image is the combination of a low frequency image and a high frequency image. First introduced by Oliva, Torralba, and Schyns in the SIGGRAPH 2006 paper. A charactor of the hybrid image is that the image looks different from different distance. This is because the one can only see low frequency part of the image from distance. In this project, I will implement a simplified version of the Hybrid Image.
Algorithms
To obtain a Hybrid Image, we need one low frequency image and one high frequency image. So we need to implement an image filter to filter out high or low frequency.Image filtering is achieved by convoluting the image with a filter. In this project, I seperately processed three channels of an color image so that the filter works both for gray scale image and color image.
Low frequency image is obtained by applying a gaussian lowpass filter.Gaussian filter is better than box filter because it's more smooth so all frequencies greater than a certain cutoff frequency are down to zero. The easiest way to get a high frequency image is to subtract a blurred version image with the original image. This method will have a similar result of applying a laplacian filter. The reason to use above method instead of using laplacian filter is that the cutoff frequency of gaussian low passfilter is the same as the blurred image. So there will be no frequency discontinue in the hybrid image.
After having a low frequnecy and a high frequency image, the next step is to merge them together. For simplicity, I just add two image together then divide the new image by two. This mtehod have pretty result, but there is also a problem that the color of hybrid image is darker than original low frequency image. Because a image is mostlty consist of low frequency, so the value of high frequency image is centered around zero, which is black. So after averaging the pixel value of the two image, the value of hybrid image is lower than the low frequency image. One possible solution to this problem is to perform a histogram equalization on the hybrid image.
Results
I created hybrid image from five image pairs given. The result of them is satisfying. But there is a need to change the cutoff frequency for each pair of images. Because some low frequency image has more edges, so I need to blur it more so that the image will not be seen in the hybrid image from a short distance. Different choice of which image to be low frequency image will also effect the outcome. In general, the main objecy in low frequency should not be smaller than the object in high frequency. Otherwise it will make the low frequency image hard to distinguish.
