Citation:
Tom Y. Ouyang and Randall Davis. 2009. A visual approach to sketched symbol recognition. In Proceedings of the 21st international jont conference on Artifical intelligence (IJCAI'09), Hiroaki Kitano (Ed.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1463-1468.
Publication Link
Summary:
The main approach followed in this paper is to represent symbols as feature images rather than temporally ordered points or geometric primitives. There are four feature images concerned with orientation, and 1 concerned with endpoint. The feature images are first extracted from the input and then there are stages of smoothing, down-sampling and classification.
Discussion:
The online stroke sequences are converted into low level feature images.
In the smoothing and downsampling phases, we try to increase the tolerance to local shifts and distortions by applying a Gaussian smoothing function. We then downsample the images by a factor of 2 using a MAX filter.
The symbol recognition is done using a deformable template matching algorithm. The IDM allows every point in the input image to shift within a 3x3 local window to find the perfect match. To compute the IDM distance, the following equation is used in the paper.
The system also incorporates two optimization schemes in the form of coarse candidate pruning (using the distance between these reduced feature sets), and hierarchical clustering (which uses a branch and bound technique, where clusters are first formed using agglomerative clustering).
Tom Y. Ouyang and Randall Davis. 2009. A visual approach to sketched symbol recognition. In Proceedings of the 21st international jont conference on Artifical intelligence (IJCAI'09), Hiroaki Kitano (Ed.). Morgan Kaufmann Publishers Inc., San Francisco, CA, USA, 1463-1468.
Publication Link
Summary:
The main approach followed in this paper is to represent symbols as feature images rather than temporally ordered points or geometric primitives. There are four feature images concerned with orientation, and 1 concerned with endpoint. The feature images are first extracted from the input and then there are stages of smoothing, down-sampling and classification.
Discussion:
The online stroke sequences are converted into low level feature images.
In the smoothing and downsampling phases, we try to increase the tolerance to local shifts and distortions by applying a Gaussian smoothing function. We then downsample the images by a factor of 2 using a MAX filter.
The symbol recognition is done using a deformable template matching algorithm. The IDM allows every point in the input image to shift within a 3x3 local window to find the perfect match. To compute the IDM distance, the following equation is used in the paper.
The system also incorporates two optimization schemes in the form of coarse candidate pruning (using the distance between these reduced feature sets), and hierarchical clustering (which uses a branch and bound technique, where clusters are first formed using agglomerative clustering).
No comments:
Post a Comment