Backpropagation Applied to Handwritten Zip Code

Comments

Andrew's blog

Summary

This paper discusses the use of neural networks in sketch recognition. In this paper the author tackles the problem of recognizing zip codes written on letters which will help in the sorting of the letters. The zip codes are written with pens on the letter paper. The system first scans the digits and then does some pre-processing on the image to retrieve the zip code in digital format.

The digits are then separated from each other and then each digits is fed into the the neural networks to be recognized. The system uses a three tier neural network layers name H1, H2 and H3. In the first layer the image is divided into pieces of 5x5 pixels and the features are extracted from the sub-parts and fed to next layer. H1 consists of 12x64 hidden units, H2 consists of 12x64 hidden units and H3 consists of 30 hidden units. The output can be mapped to 10 units which are the digits in the number series.

According to the author the system misclassified patterns was 0.14% on training set and 5% on testing set.

Discussion

This paper discusses an alternate research area of sketch recognition. Neural networks can also be effectively used in the process of classifying the input sketch to an output sketch. I think the paper lacked detailed information on the what features the author used to train the neural networks.

Also his technique was tested on a very small domain in which there is not high variation in the input data.