CMPT414 Couse project proposal

I'm currently taking CMPT414 model based computational vision. We need to do a final term project. One of my classmates and I formed a team and proposed a project about a neural network specifically for recognizing handwritten digits. Ideally, the final product would take images of handwritten digits and output predicated digit, ranging from [0, 9], with comparatively high accuracy. The product will have various real world applications. One example would be a Chinese postal code reader as in China the ZIP code consists of 5 numeric digits. If more time is available, we would also like to expand our project to real world data. For example, we can use some scanned real handwritten images as the input of the built neural network and see the results. This will involve some image pre-processing techniques, such as noise removal, skew detection and correction, character normalization and so on.
We will mainly focus on the implementation of the algorithm and leave out the UI part. Matlab will be our main programming tool.

MNIST database (http://yann.lecun.com/exdb/mnist/index.html ) will be used as the benchmark for this project. The MNIST database of handwritten digits consists of a training set of 60,000 examples, and a test set of 10,000 examples. It has been widely used in lots of projects or research works. Dr. Yann LeCun (http://yann.lecun.com/) has implemented a neural network which achieves 99.18% accuracy (i.e., an error rate of only 0.82%). This error rate served as a type of "benchmark" for many others’ research works and could be the goal of our project as well, which might be challenging. The data files of this database are not in any standard image format. We will have to write our own program to read them.

Mike O'Neill has implemented a complete project for Handwritten Digits Recognition, with demonstration graphic UI. He has released the details of his design in the article of Neural Network for Recognition of Handwritten Digits. Mike built Five-layer Convolutional Neural Network and applied Second Order Methods in back propagation to optimize the performance. Mike’s digit recognition project achieves 99.26% accuracy on MNIST dataset. We will follow his design and implement the essential algorithm.

I think the idea and design is pretty cool. I will be very happy if we can achieve above 99% accuracy!