Progress Report

1) Plots and Analysis of Data

For plots and analysis of our data, please see the pages of our website under "Image Processing" where we apply the 2D DFT, edge detection, and experiment with other image processing techniques.

2) Work and Achievement So far

We have conducted many experiments on the satellite images over the last few weeks. We wish to find a way to adjust the resolution of satellite image to make it same as road map so we can use correlate function easily. We also want to extract the road out of the satellite image by filters. We first apply 2D FFT on the input data, and plot the coefficient. Comparing DFT image of satellite image and road map, we find that their behaviors are quite similar to each other in frequency domain. That means it could be possible to apply filters in frequency domain to correlate satellite image and road map.

Then, since road edges are parallel curves in maps, we decided to use edge detection to see if we can make a breakthrough. Among all the three edge detection method provided by Matlab, Sobel's method give best result. However, as the figure shows, the processed binary image leaves a lot of noise which make it almost impossible to correlate with road map.

We also tried other process method on the satellite image. Since edge detection is based on gradient of digital image, we tried to calculate the gradient and show the result graphically. The result is shown in "Image processing".

At last, we experiment color threshold method to extract the road out. The result shown is satisfying, the road is shown clearly and the noise is tiny, which means we color threshold could be a potential solution of satellite image processing.

For image correlation, we found the Matlab function normcorr2 useful. The function calculate 2D correlation of one image and its template, and the position with the highest value should be the position where the image matches. We tried to correlate a image and a part cropped from the image and success. The experiment images are shown in "Correlation Algorithm"

3) Plan

From the previous work, it is shown that we still need a better filter to extract roads on satellite image with less noise. Also, the scaling is a problem not solved. We haven't experiment any filter with this function yet.

We also need improve correlation algorithm since the processed satellite image has a much worse quality than an image cropped from original road map.

Important tasks in order of importance:

Apply the current correlation algorithm to our image filtering output. Identify the limitations of the approach and generate improvements.
Select the image filtering output that will give us the best results in later steps. Improve and adjust the image filtering as necessary.
Write a driver function that can succinctly bring together a demonstration of our project.

4) Summary

To succeed in our project, we have studied many filters and DSP operation that is not mentioned in class. For example, the color threshold filter and 2D matrix correlation.

An interesting topic we have explored and learned about is the re-sampling of a signal by modifications in the frequency domain. This is important to our project because the correlation algorithm will require that the satellite image is of the same size as the feature in the road map with which it should match. Therefore, the images must have the same resolution. We have shown that by using simple operations on the DFT such as cropping out the high frequencies we can downsample an image in a smooth way. When doing this, we noticed that the intensity of our image significantly increased. This is because MATLAB does normalization only in the inverse transform which does not reduce magnitude enough. We therefore add our own scaling after the fact. We feel that is topic is a good blend of interesting, applicable information that very clearly builds upon the theory learned in class.