Edge Detection - Kayle Gishen

Overview

Edge detection algorithms are generally succeptible to texture information, however, if we use texture information to remove these false positives the contour of the object can be identified more easily.

Algorithm

The idea of pb-lite is to use Canny edge detectors to find all posible edges in an image then use texture and brightness gradients to determine if the edge is correct. The texture information is stored in a texton map, which is generated by filtering th image with different filters and then using k-means to group the features.

Filters

Several different filters where tried, including: Gaussian, Gaussian * Sobel, and Gabor. In the Gaussian cases, the sigma sizes where adjusted, and the bandwidth was adjusted for the Gabor filters. In each cases 32 filters were used.

Gabor Filterbank

Gaussian Filterbank

Gaussian * Sobel Filterbank

Masks

For all trials the masks consisted of binary half-disk pairs with radii of 5, 10, and 20.

Results

Using the provided benchmark, the different methods are compared using their resulting f-score.

From the values above we can see that the Gabor filters had the best response in terms of F-Score, but the plain Gaussian had the best Precision/Recall Area. Therefore, the results seem to show that orientation information is not as dominant as the center surround response of tha Gaussian.

Sample Images : Gaussian - Gaussian * Sobel - Gabor pb - Gabor Energy

Additional Work

Following the use of the Gabor filters, the Gabor energy function was implemeted as well. The Gabor energy function performed poorly in the benchmark since it does not classify most background (soft focus) edges as edges which the benchmark uses. However, the images resulting from the Gabor energy function were compared to the Gaussian pb-lite results by several human subjects, with each stating that the Gabor results are better represetations of the subject of the image. Since the Gabor filters model the simple and complex cells used by humans in visual perception, it makes sense that the results are better understood by humans.

Conclusions

The pb-lite concept seems to be a successful edge detector when attempting to classify a scene and segmenting an image. The slow running time of the algorithm prevents it from being useful in realtime applications however. For detecting an object in an image, the Gabor system works very quickly and forms a very accurate outline of the object and still maintains subtle depth information in the image. pb-lite is based mainly on the canny edge detector, so achieving better results than canny is difficult since canny provides the most influence to pb-lite.