Bibliography:
Levent Burak Kara and Thomas F. Stahovich. An Image-based, Trainable Symbol Recognizer for Hand-drawn Sketches. Computers and Graphics. Volume 29, Issue 4. August, 2005.
Summary:
The paper talks about an image based trainable, multi-stroke hand drawn symbol recognizer. Some advantages of this image based recognizer is that it avoids the segmentation and feature extration problems, commonly seen in recognizers based on single-stroke recognizers that break an image into constituent primitive lines and curves.
The recognizer can learn new symbol definitions from just one prototype image, allowing users to train the system with new symbols on the fly. The system mainly consists of two steps: 1) Make the image rotation invariant. This is done by transforming the image into polar coordinates and match with given definitions using rotation. Additionally, dissimilar images can be pruned off at this stage. In the next step, multiple classifiers analyze the remaining candidates and produce a list of definitions ranked by similarity using template matching. The template matching is done using Modified Hausdorff distance, Tanimoto similarity coefficient and the Yule coefficient. Finally, the individual classification results are normalized and combined to make the final decision.
A user study was conducted and it was found that the system accurately classifies symbols with only a small amount of training data and works well in both user-dependent and independent studies.
Discussion:
This is the first time I am reading about a recognizer that tries to identify the entire image rather than its constituent primitives. While this works great for specific applications and as an independent system, I wonder if it can be combined with other systems to form a higher level system. That way, we can use the output of this system as a feature for another system.
Levent Burak Kara and Thomas F. Stahovich. An Image-based, Trainable Symbol Recognizer for Hand-drawn Sketches. Computers and Graphics. Volume 29, Issue 4. August, 2005.
Summary:
The paper talks about an image based trainable, multi-stroke hand drawn symbol recognizer. Some advantages of this image based recognizer is that it avoids the segmentation and feature extration problems, commonly seen in recognizers based on single-stroke recognizers that break an image into constituent primitive lines and curves.
The recognizer can learn new symbol definitions from just one prototype image, allowing users to train the system with new symbols on the fly. The system mainly consists of two steps: 1) Make the image rotation invariant. This is done by transforming the image into polar coordinates and match with given definitions using rotation. Additionally, dissimilar images can be pruned off at this stage. In the next step, multiple classifiers analyze the remaining candidates and produce a list of definitions ranked by similarity using template matching. The template matching is done using Modified Hausdorff distance, Tanimoto similarity coefficient and the Yule coefficient. Finally, the individual classification results are normalized and combined to make the final decision.
A user study was conducted and it was found that the system accurately classifies symbols with only a small amount of training data and works well in both user-dependent and independent studies.
Discussion:
This is the first time I am reading about a recognizer that tries to identify the entire image rather than its constituent primitives. While this works great for specific applications and as an independent system, I wonder if it can be combined with other systems to form a higher level system. That way, we can use the output of this system as a feature for another system.
No comments:
Post a Comment