Proceedings of the 17th International Conference on Pattern Recognition, 2004. ICPR 2004. 2004
DOI: 10.1109/icpr.2004.1333823
|View full text |Cite
|
Sign up to set email alerts
|

Local context in non-linear deformation models for handwritten character recognition

Abstract: We evaluate different two-dimensional non-linear deformation models for handwritten character recognition. Starting from a true two-dimensional model, we derive pseudo-two-dimensional and zero-order deformation models. Experiments show that it is most important to include suitable representations of the local image context of each pixel to increase performance. With these methods, we achieve very competitive results across five different tasks, in particular 0.5% error rate on the MNIST task.

Help me understand this report

Search citation statements

Order By: Relevance

Paper Sections

Select...
2
2
1

Citation Types

1
26
0

Year Published

2005
2005
2007
2007

Publication Types

Select...
6
2

Relationship

0
8

Authors

Journals

citations
Cited by 31 publications
(27 citation statements)
references
References 13 publications
(7 reference statements)
1
26
0
Order By: Relevance
“…In order to extract a silhouette for each digit, we considered only the pixels with intensity values higher than 0.8. The average error rate obtained in 20 random experiments was 2:1% AE 0:3%, which is comparable to the state-of-the-art results (1.9 percent) reported on this database in [29].…”
Section: Usps Databasesupporting
confidence: 74%
See 1 more Smart Citation
“…In order to extract a silhouette for each digit, we considered only the pixels with intensity values higher than 0.8. The average error rate obtained in 20 random experiments was 2:1% AE 0:3%, which is comparable to the state-of-the-art results (1.9 percent) reported on this database in [29].…”
Section: Usps Databasesupporting
confidence: 74%
“…The average error rate obtained in 20 random experiments was 2% AE 0:3%. Our results are superior to many of the algorithms tested in the benchmark [34], including RBF, Neural nets, PCA, and linear classifiers, but are inferior to the best results published for boosted LeNet-4 [34] (0.7 percent), shape context [5] (0.63 percent), virtual SVM [15] (0.56 percent), local context and nonlinear deformation models [29] (0.43 percent), and convolutional neural networks [51] (0.4 percent). However, our algorithm is fast.…”
Section: Mnist Databasementioning
confidence: 83%
“…To compare images, the Euclidean distance can be seen as a very basic baseline, and in earlier works it was shown that image deformation models are a suitable way to improve classification performance significantly e.g. for medical radiographs and for optical character recognition [11,10]. Here we allow each pixel of the database images to be aligned to the pixels from a 5×5 neighborhood from the image to be classified taking into account the local context from a 3×3 Sobel neighborhood.…”
Section: Image Distortion Modelmentioning
confidence: 99%
“…The image distortion model [10,11] is a zeroth-order image deformation model to compare images pixel-wise. Here, classification is done using the nearest neighbor decision rule: to classify an image, it is compared to all training images in the database and the class of the most similar image is chosen.…”
Section: Image Distortion Modelmentioning
confidence: 99%
“…In particular this notion of invariant distance covers many specific examples in literature. In particular tangent distance (TD) (Simard et al 1993), the image distortion model (IDM) (Keysers et al 2000), general deformation models (Keysers et al 2004), dynamic time warping (DTW) (Rabiner and Juang 1993), Fréchet distance (Alt and Guibas 1999), invariant distances between point sets (Werman and Weinshall 1995) or two-sided manifold distance (Fitzgibbon and Zisserman 2003).…”
Section: Invariant Distance Substitution Kernelsmentioning
confidence: 99%