Paper Results

I’ve completed my paper and is in the process of submission. Hopefully the paper will pass through the peer review unscathed, but I suppose few are that lucky.

Since it will be a while before the paper is in print, I shall give a sneak informal preview of the results.

In the paper, we investigated two matters:

1. Image processing is typically done on a pixel-by-pixel basis. Recently, a number of approaches have opted to segment an image and process each segment as a whole. We investigated whether there is any gain in accuracy when adding a segmentation component into our road extraction system.
2. Some road extraction systems employs several classifier to identify road pixels/segments. The outputs from these classifiers is then fused in order to provide more consistent and accurate results. Again we investigate whether this holds for our system.

The input image and ground truth:

Input Image	Ground Truth

1. Segmentation vs. pixel based approaches:

Here is the results from the segment based approach:

Classified Image	Thresholded Result	SORM Output

Here is the results from the pixel based approach:

Classified Image	Thresholded Result	SORM Output

The outcome was that in this case (using the fusion of classifiers) the segmented approach (S) provided better results than the pixel based approach (P). S achieved a quality measurement of 60.7% whilst P managed 51.5%. Furthermore, the standard deviation for S was lower at 15.4% than P’s 18.8%.

2. Single classifier vs. data fusion for pixel and segmented approaches:
For the second experiment, we removed the fusion component and tested each classifier separately using the pixel and segmented approaches.

For P, the classifier using the Mahalanobis distance function achieved the highest quality. Here is some sample images:

Classified Image	Thresholded Result	SORM Output *

For S, the classifier using the Bhattacharrya distance function achieved the highest quality. Here is some sample images:

Classified Image	Thresholded Result	SORM Output *

* Please note that the images above for the Mahalanobis pixel based approach (PM) and Bhattacharyya segment based approach (SB) might not be representative of the entire data set we used.

The results showed that PM achieved an quality of 64.6% with a 17.3% standard deviation, opposed to SB’s 62.8% and 13.6%.

One can therefore not conclude that S will always have a higher quality measurement, but does definitely prove to be more consistent than P.

Furthermore, data fusion reduced the quality of our results and we suspect that it is because the problem only has two classes (road and non-road). The classifiers are constantly in direct conflict/agreement of the result. We suspect data fusion to be more suited for applications where say sensor 1 detected an object as being A or B and sensor 2 detected and the same object as being B or C. The fusion of this result should yield the correct answer of B.

Our next step would be to create a higher level system to breach the gaps that exist in the results. We hope to raise our quality measurements above 70%.

~ by ahauptfleisch on November 21, 2007.